research paper on sentiment analysis using machine learning

• Open access
• Published: 11 December 2017

Optimization of sentiment analysis using machine learning classifiers

• Jaspreet Singh   ORCID: orcid.org/0000-0002-9018-1233 1 ,
• Gurvinder Singh 1 &
• Rajinder Singh 1  

Human-centric Computing and Information Sciences volume  7 , Article number:  32 ( 2017 ) Cite this article

34k Accesses

104 Citations

2 Altmetric

Metrics details

Words and phrases bespeak the perspectives of people about products, services, governments and events on social media. Extricating positive or negative polarities from social media text denominates task of sentiment analysis in the field of natural language processing. The exponential growth of demands for business organizations and governments, impel researchers to accomplish their research in sentiment analysis. This paper leverages four state-of-the-art machine learning classifiers viz. Naïve Bayes, J48, BFTree and OneR for optimization of sentiment analysis. The experiments are performed using three manually compiled datasets; two of them are captured from Amazon and one dataset is assembled from IMDB movie reviews. The efficacies of these four classification techniques are examined and compared. The Naïve Bayes found to be quite fast in learning whereas OneR seems more promising in generating the accuracy of 91.3% in precision, 97% in F-measure and 92.34% in correctly classified instances.

Introduction to sentiment analysis

The popularity of rapidly growing online social networks and electronic media based societies has influenced the young researchers to pursue their work on sentiment analysis. These days organizations quite keen assess their customers or public opinion about their products from social media text [ 1 ]. The online service providers are hooked on assessing social media data on blogs, online forums, comments, tweets and product reviews. This assessment is exploited for their decision making or amelioration of their services or quality of products. The applications of sentiment analysis encompass the areas like social event planning, election campaigning, healthcare monitoring, consumer products and awareness services. The immoderate use of internet by business organizations all around the globe has noticed that opinionated web text has molded our business plays and socio-economic systems. The computational power is fueled by burgeon of machine learning techniques. This work focused on four text classifiers utilized for sentiment analysis viz. Naïve Bayes, J48, BFTree and OneR algorithm. The “ Machine learning techniques for sentiment analysis ” section of this paper provides the intuition behind the task of sentiment classification by leveraging the modeling of aforementioned four classifiers. The architecture of proposed model using four sentiment classifiers is disposed in “ Proposed methodology for optimization of sentiment prediction using weka ” section. The related work with recent contributions of machine learning in the field of sentiment classification is described in “ Related work ” section. In “ Datasets taken ” section, the three manually annotated datasets are described along with their preprocessing. The experimental results and discussion of efficacies of classifiers are cataloged in “ Results and discussions ” section followed by the ending remarks along with a future direction in “ Conclusion ” section.

Levels of sentiment

Due to scarcity of opinion text available in digital form, very less research interest on computational linguistics in the last decade of twentieth century was witnessed [ 2 , 3 , 4 ]. The escalation of social media text on internet attracts young researchers to define the level of granularities of text. The web text is classified into three levels viz. document level, sentence level and word level. In [ 5 ], the fourth level granularity is defined by using deep convolution neural network. This fourth level is character level feature extraction approach used for extracting features of each character window from given word (Table  1 ).

Machine learning techniques for sentiment analysis

The social networking sites dispense their data conveniently and freely on the web. This availability of data entices the interest of young researchers to plunge them in the field of sentiment analysis. People express their emotions and perspectives on the social media discussion forums [ 6 ]. The business organizations employ researchers to investigate the unrevealed facts about their products and services. Spontaneous and automatic determination of sentiments from reviews is the main concern of multinational organizations [ 7 , 8 , 9 , 10 ]. The machine learning techniques have improved accuracy of sentiment analysis and expedite automatic evaluation of data these days. This work attempted to utilize four machine learning techniques for the task of sentiment analysis. The modeling of four techniques is briefly discussed below.

Naïve Bayes used for sentiment classification

The dichotomy of sentiment is generally decided by the mindset of an author of text whether he is positively or negatively oriented towards his saying [ 6 , 11 , 12 , 13 ]. Naïve Bayes classifier is a popular supervised classifier, furnishes a way to express positive, negative and neutral feelings in the web text. Naïve Bayes classifier utilizes conditional probability to classify words into their respective categories. The benefit of using Naïve Bayes on text classification is that it needs small dataset for training. The raw data from web undergoes preprocessing, removal of numeric, foreign words, html tags and special symbols yielding the set of words. The tagging of words with labels of positive, negative and neutral tags is manually performed by human experts. This preprocessing produces word-category pairs for training set. Consider a word ‘y’ from test set (unlabeled word set) and a window of n-words (x 1 , x 2 , …… x n ) from a document. The conditional probability of given data point ‘y’ to be in the category of n-words from training set is given by:

Consider an example of a movie review for movie “Exposed”. The experimentation with Naïve Bayes yields the following results.

J48 algorithm used for sentiment prediction

The hierarchical mechanism divides feature space into distinct regions followed by the categorization of sample into category labels. J48 is a decision tree based classifier used to generate rules for the prediction of target terms. It has an ability to deal with larger training datasets than other classifiers [ 14 ]. The word features for sentences of corpus taken from labeled arff file of training set are represented in the leaf nodes of decision tree. In the test set every time when a near feature qualifies the label condition of internal feature node, its level is lifted up in the same branch of decision tree. The assignment of labels to the word features of test set gradually generates different two branches of decision tree. J48 algorithm uses entropy function for testing the classification of terms from the test set.

where (Term) can be unigram, bigram and trigram. In this study we have considered unigrams and bigrams. The example in the Table  2 contains bigrams like “Horrible acting”, “Bad writing” and “Very misleading” are labeled with negative sentiment whereas the term “More enjoyable” reflects positive sentiment towards the movie. The decision tree of J48 algorithm for obtaining sentiment form text is represented in the Fig.  1 below.

J48’s Decision Tree for terms of Example in Table  2

BFTREE algorithm used for sentiment prediction

Another classification approach outperforms J48, C4.5 and CART by expanding only best node in the depth first order. BFTree algorithm excavates the training file for locating best supporting matches of positive and negative terms in the test file. BFTree algorithm keeps heuristic information gain to identify best node by probing all collected word features. The only difference between J48 and BFTree classifier is the computation order in which decision tree is built. The decision tree disparate feature terms of plain text taken from movie reviews and classify them at document level by tagging appropriate labels. BFTree extracts best node from labeled and trained binary tree nodes to reduce the error computed from information gain.

where S is word feature term of test set and A is the attribute of sampled term from training set. V(A) denotes set of all possible values of A. The binary tree stops growing when an attribute A captures single value or when value of information gain vanishes.

OneR algorithm used for sentiment prediction

OneR algorithm is a classification approach which restricts decision tree to level one thereby generating one rule. One rule makes prediction on word feature terms with minimal error rate due to repetitive assessment of word occurrences. The classification of most frequent terms of a particular sentence is made on the basis of class of featured terms from training set. The demonstration of OneR algorithm for sentiment prediction with smallest error of classification is given below:

Select a featured term from training set.

Train a model using step 3 and step 4.

For each prediction term.

For each value of that predictor.

Count frequency of each value of target term.

Find most frequent class.

Make a rule and assign that class to predictor.

Calculate total error of rules of each predictor.

Choose predictor with smallest error.

Proposed methodology for optimization of sentiment prediction using weka

The preprocessing of raw text from web is done in python 3.5 using NLTK and bs4 libraries. Each review in the first dataset is parsed with NLTK’s parser and title of the review is considered as a feature. We have obtained 15 features from first dataset and 42 features from each of second and third dataset. The CSV files generated from Python are converted to ARFF files for WEKA 3.8. Only two sentiment labels namely Pos for positive and Neg for negative are used for assigning sentences. The working methodology of proposed work for optimization of sentiment prediction is given below in Fig.  2 .

Proposed methodology

After loading files with ARFF loader, the class assigner picks up appropriate class labels from dataset and performs feature selection on the basis of frequently used headings and most frequent titles. The feature selector module is implemented using three feature selection methods namely Document Frequency (DF), Mutual Information (MI) and Information Gain (IG). The mathematical modeling of these feature selection methods requires some probability distributions and statistical notations described below:

P(w): Probability that a document ‘d’ contains term ‘w’.

P(c’): Probability that document ‘d’ does not belongs to category ‘c’.

P(w, c): Joint probability that document ‘d’ contains word term ‘w’ of category ‘c’.

P(c/w): Conditional probability that a document ‘d’ belongs to category ‘c’ under the condition that ‘d’ contains word term ‘w’.

Similarly other notations like P(w’), P(w/c), P(w/c’), P(c/w’) and P(c’/w) are taken and {c} is the set of categories.

N 1 : Number of documents that exhibit category ‘c’ and contain term ‘w’.

N 2 : Number of documents that do not belong to category ‘c’ but contains term ‘w’.

N 3 : Number of documents that belong to category ‘c’ and do not contain term ‘w’.

N 4 : Number of documents that neither belong to category ‘c’ nor contain term ‘w’.

N: Total number of document reviews.

DF method qualifies only those documents in which a higher frequency terms are considered.

The MI method measures features of text by computing similarity of word terms ‘w’ and category ‘c’.

The IG-construct measures similarity information for category by exploiting probabilities of absence or presence of terms in a document review.

The normalization module converts all letters into lowercase, removal of punctuation marks and special symbols, conversion of numbers into words, expansion of abbreviation and limiting the average length of twenty words in a sentence. Each sentence is delimited by a newline character. The Python’s NLTK and bs4 libraries are used for this purpose. Data splitter take the ratio of (80:20) of (Train: Test) subsets. We have used manual splitting of dataset at the time of retrieval of data from web. The four classifiers are trained with training subsets followed by performance evaluation. The evaluation metrics taken in the experiment are precision, recall, accuracy and F-measure.

Related work

Existing approaches of sentiment prediction and optimization widely includes SVM and Naïve Bayes classifiers. Hierarchical machine learning approaches yields moderate performance in classification tasks whereas SVM and Multinomial Naïve Bayes are proved better in terms of accuracy and optimization. Sentiment analysis using neural network architectures has appeared in very few works. The sentiment prediction methods using recursive neural networks and deep convolution neural networks are bit complex in capturing compositionality of words. Extracting character level features and embeddings of complex words is found hard in many neural network architectures whereas extracting sentence level or word level features such as morphological tags and stems are more effectively achieved in convolutional neural networks. A very few researchers have used J48, BFTree and OneR for the task of sentiment prediction. These three classifiers are utilized for other classification tasks like emotion recognition from text and twitter’s text categorizations. The summary of benchmarks related to machine learning techniques in terms of accuracy of classification is listed in the Table  2 . SVM and Naive Bayes are proving better in terms of benchmarks than other machine learning techniques (Table  3 ).

Datasets taken

Three Datasets are manually annotated from http://www.amazon.in . First dataset consists of product reviews of Woodland’s wallet are taken from 12th October 2016 to 25th October 2016 for training set containing 88 reviews and from 25th October 2016 to 30th October 2016 for testing set containing 12 randomly chosen product reviews with their sentiments prediction using four machine learning algorithms. Second dataset consists of 7465 Digital Camera reviews of Sony are taken from 01st October 2016 to 25th October 2016 for training set and 1000 reviews are from 25th October 2016 to 30th October 2016 for test dataset. Third dataset consists of movie reviews taken from http://www.imdb.com . It contains 2421 reviews for training set and 500 reviews for test set.

Results and discussions

The experiment is carried out by using freeware WEKA software tool for classification of sentiments in the text. Standard implementations of Naïve Bayes, J48, BFTree and OneR algorithms are exploited from WEKA version 3.8. The classification accuracy of first dataset shows 100% classification accuracy with Naïve Bayes in some of the epochs because of small size of dataset. The average of 29 epochs for all four classifiers on second and third datasets is presented in Table  4 below. Naïve Bayes shows faster learning among four classifiers whereas J48 found to be slower. OneR classifier is leading from other three classifiers in percentage of correctly classified instances. The accuracy of J48 algorithm is promising in true positive and false positive rates.

Results of classification accuracies for the test subsets with 42 and 15 attributes are recorded. The average accuracies of 29 runs on three datasets is presented in Table  5 below. All four classifiers improved in accuracies with the increase of features from 15 to 42. This shows the direct proportionality of multiple features with learning capability for machine learning algorithms.

This paper exploits four machine learning classifiers for sentiment analysis using three manually annotated datasets. The mean of 29 epochs of experimentation recorded in Table  4 shows that OneR is more precise in terms of percentage of correctly classified instances. On the other hand, Naïve Bayes exhibits faster learning rate and J48 reveals adequacy in the true positive and false positive rates. Table  5 reveals the truth that J48 and OneR are better for smaller dataset of woodland’s wallet reviews. The preprocessing of proposed methodology is limited to extract foreign words, emoticons and elongated words with their appropriate sentiments. The future work in the task of sentiment analysis has scope to improve preprocessing with word embeddings using deep neural networks and can also extend this study through convolution neural networks.

Parvathy G, Bindhu JS (2016) A probabilistic generative model for mining cybercriminal network from online social media: a review. Int J Comput Appl 134(14):1–4. doi: 10.5120/ijca2016908121

Google Scholar  

Cambria E, White B (2014) Jumping NLP curves: a review of natural language processing research. IEEE Comput Intell Mag 9(2):48–57. doi: 10.1109/mci.2014.2307227

Article   Google Scholar  

Pang B, Lee L, Vaithyanathan S (2002) Thumbs up? In: Proceedings of the ACL-02 conference on empirical methods in natural language processing—EMNLP ‘02. doi: 10.3115/1118693.1118704

Poria S, Cambria E, Gelbukh A, Bisio F, Hussain A (2015) Sentiment data flow analysis by means of dynamic linguistic patterns. IEEE Comput Intell Mag 10(4):26–36. doi: 10.1109/mci.2015.2471215

Nogueira dos Santos C, Gatti M (2014) Deep convolution neural networks for sentiment analysis of short texts. In: Proceedings of COLING 2014, the 25th international conference on computational linguistics. p 69–78

Kiritchenko S, Mohammad S, Salameh M (2016) SemEval-2016 task 7: determining sentiment intensity of English and Arabic phrases. In: Proceedings of the 10th international workshop on semantic evaluation (SemEval-2016). doi:10.18653/v1/s16-1004

Cernian A, Sgarciu V, Martin B (2015) Sentiment analysis from product reviews using SentiWordNet as lexical resource. In: 2015 7th international conference on electronics, computers and artificial intelligence (ECAI). doi: 10.1109/ecai.2015.7301224

Hammer HL, Solberg PE, Øvrelid L (2014) Sentiment classification of online political discussions: a comparison of a word-based and dependency-based method. In: Proceedings of the 5th workshop on computational approaches to subjectivity, sentiment and social media analysis. doi: 10.3115/v1/w14-2616

Zadeh L (2006) Toward human-level machine intelligence. In: 2006 18th IEEE international conference on tools with artificial intelligence (ICTAI’06). doi: 10.1109/ictai.2006.114

Joachims T (2002) Text classification. Learning to classify text using support vector machines. p 7-33. doi: 10.1007/978-1-4615-0907-3_2

Wanton TM, Porrata AP, Guijarro AM, Balahur A (2010) Opinion polarity detection—using word sense disambiguation to determine the polarity of opinions. In: Proceedings of the 2nd international conference on agents and artificial intelligence. doi: 10.5220/0002703504830486

Xia Y, Cambria E, Hussain A, Zhao H (2014) Word polarity disambiguation using bayesian model and opinion-level features. Cogn Comput 7(3):369–380. doi: 10.1007/s12559-014-9298-4

Dey L, Chakraborty S, Biswas A, Bose B, Tiwari S (2016) Sentiment analysis of review datasets using Naïve Bayes’ and K-NN classifier. Int J Inform Eng Electron Bus 8(4):54–62. doi: 10.5815/ijieeb.2016.04.07

Nie CY, Wang J, He F, Sato R (2015) Application of J48 decision tree classifier in emotion recognition based on chaos characteristics. In: Proceedings of the 2015 international conference on automation, mechanical control and computational engineering. doi: 10.2991/amcce-15.2015.330

Tan S, Zhang J (2008) An empirical study of sentiment analysis for Chinese documents. Expert Syst Appl 34(4):2622–2629. doi: 10.1016/j.eswa.2007.05.028

Mohammad SM, Zhu X, Kiritchenko S, Martin J (2015) Sentiment, emotion, purpose, and style in electoral tweets. Inf Process Manage 51(4):480–499. doi: 10.1016/j.ipm.2014.09.003

Kiritchenko S, Mohammad SM (2016) sentiment composition of words with opposing polarities. In: Proceedings of the 2016 conference of the north american chapter of the association for computational linguistics: human language technologies. doi:10.18653/v1/n16-1128

Dashtipour K, Poria S, Hussain A, Cambria E, Hawalah AY, Gelbukh A, Zhou Q (2016) Multilingual sentiment analysis: state of the art and independent comparison of techniques. Cogn Comput 8(4):757–771. doi: 10.1007/s12559-016-9415-7

Sobhani P, Mohammad S, Kiritchenko S (2016) Detecting stance in tweets and analyzing its interaction with sentiment. In: Proceedings of the 5th joint conference on lexical and computational semantics. doi:10.18653/v1/s16-2021

Poria S, Cambria E, Winterstein G, Huang G (2014) Sentic patterns: dependency-based rules for concept-level sentiment analysis. Knowl Based Syst 69:45–63. doi: 10.1016/j.knosys.2014.05.005

Socher R (2016) deep learning for sentiment analysis—invited talk. In: Proceedings of the 7th workshop on computational approaches to subjectivity, sentiment and social media analysis. doi:10.18653/v1/w16-0408

Turney PD, Mohammad SM (2014) Experiments with three approaches to recognizing lexical entailment. Nat Lang Eng 21(03):437–476. doi: 10.1017/s1351324913000387

Mohammad S, Kiritchenko S, Sobhani P, Zhu X, Cherry C (2016) SemEval-2016 task 6: detecting stance in tweets. In: Proceedings of the 10th international workshop on semantic evaluation (SemEval-2016). doi:10.18653/v1/s16-1003

Poria S, Chaturvedi I, Cambria E, Hussain A (2016) Convolutional MKL based multimodal emotion recognition and sentiment analysis. In: 2016 IEEE 16th international conference on data mining (ICDM). doi: 10.1109/icdm.2016.0055

Wang J, Yu L, Lai KR, Zhang X (2016) Dimensional sentiment analysis using a regional CNN-LSTM model. In: Proceedings of the 54th annual meeting of the association for computational linguistics, vol 2: short papers. doi:10.18653/v1/p16-2037

Download references

Authors’ contributions

JS made substantial contributions to conception and design, or acquisition of data, or analysis and interpretation of data. GS helped in revision and has given final approval of the version to be published. RS agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. All authors read and approved the final manuscript.

Acknowledgements

This research was supported by Department of Computer Science, Guru Nanak Dev University, Amritsar. I thank Dr. Gurvinder Singh and Dr. Rajinder Singh for their participation in experimental work and their assistance to improve the manuscript.

Competing interests

This research work has non-financial Academic and intellectual competing interests.

Availability of data and materials

I submit that I can make the experimental data and materials available after the completion of my thesis.

Consent for publication

We hereby grant and assign all rights to Human-centric Computing and Information Sciences for publication.

Ethics approval and consent to participate

The article submitted is an original work and has neither been published in any other peer-reviewed journal nor under consideration for publication by any other journal.

Funding information

No funding was received from any funder.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Author information

Authors and affiliations.

Department of Computer Science, Guru Nanak Dev University, Amritsar, India

Jaspreet Singh, Gurvinder Singh & Rajinder Singh

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Jaspreet Singh .

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License ( http://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and permissions

About this article

Cite this article.

Singh, J., Singh, G. & Singh, R. Optimization of sentiment analysis using machine learning classifiers. Hum. Cent. Comput. Inf. Sci. 7 , 32 (2017). https://doi.org/10.1186/s13673-017-0116-3

Download citation

Received : 24 December 2016

Accepted : 04 September 2017

Published : 11 December 2017

DOI : https://doi.org/10.1186/s13673-017-0116-3

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Sentiment analysis
• Social media text
• Movie reviews
• Product reviews
• Machine learning classifiers

• Open access
• Published: 23 November 2023

Sentiment analysis from textual data using multiple channels deep learning models

• Adepu Rajesh 1 , 2 &
• Tryambak Hiwarkar 1  

Journal of Electrical Systems and Information Technology volume  10 , Article number:  56 ( 2023 ) Cite this article

1092 Accesses

Metrics details

Text sentiment analysis has been of great importance over the last few years. It is being widely used to determine a person’s feelings, opinions and emotions on any topic or for someone. In recent years, convolutional neural networks (CNNs) and long short-term memory (LSTM) have been widely adopted to develop such models. CNN has shown that it can effectively extract local information between consecutive words, but it lacks in extracting contextual semantic information between words. However, LSTM is able to extract some contextual information, where it lacks in extracting local information. To counter such problems, we applied the attention mechanism in our multi-channel CNN with bidirectional LSTM model to give attention to those parts of sentence which have major influence in determining the sentiment of that sentence. Experimental results show that our multi-channel CNN model with bidirectional LSTM and attention mechanism achieved an accuracy of 94.13% which outperforms the traditional CNN, LSTM + CNN and other machine learning algorithms.

Introduction

Sentiment analysis, also called opinion mining, is the field of study that examines people’s opinions, sentiments, attitudes and emotions toward entities such as products, services, organizations, individuals, issues, events, topics and their attributes [ 1 ]. Due to wide availability of Internet and mobile networks, there is more and more textual data being published over the Internet. People are expressing their views or comments over the Internet, either on social media or on sites in the form of reviews. Every organization wants to analyze the reviews it is getting online, mostly, the product’s success depends on the reviews, generally, positive reviews. For instance, the IMDB [ 16 , 41 ] dataset contains reviews of people regarding different movies. Everyone wants to see a film, only if it has positive reviews. So, nowadays more research is being focused to analyze these reviews and predict sentiment with minimum errors.

Many sentiment analysis models use machine learning (ML) methods such as naïve Bayes and support vector machines [ 2 ]. But recently, due to increased amount of data or textual information, researchers have been greatly utilizing deep neural networks (DNNs) for this task. The CNN is basically applied for the image handling task such as mental disorder [ 37 ] and video surveillance [ 38 ]. Moreover, LSTM utilization apart from text processing can be identified in fault detection [ 39 ] in induction motor, electro-oculogram signal classification [ 40 ], etc. In deep neural networks, CNNs have emerged as very effective tool in analyzing text, especially the local information of sentence. The multi-channel CNN has provided even greater results in this specific field. But, the CNNs lack behind in extracting the semantic contextual information between words. In addition, LSTMs lack behind CNN in terms of getting the local information. To tackle these problems, we have developed a multi-channel CNN + BiLSTM model with attention mechanism; after experimentation, it is observed as the revolutionary change in text analysis, i.e., sentiment analysis task.

In this paper, we have first used the GloVe embeddings [ 3 ] to vectorize the representation of words in a meaningful way; then, this embedding is fed to multi-channel CNN with each CNN layer having separate BiLSTM and attention layer. Later, these results after all processing are passed through a multilayer perceptron network to classify the text as positive or negative. The main objective of this paper is to offer a better classifier model to predict sentiment from IMDB dataset.

The paper organization is as follows—" Related work " section highlights the literature review related to sentiment analysis. Model development is presented in section " Methodology. " The section " Experimental setup and dataset " demonstrates experimental setup, dataset and data preprocessing. Results and discussions are shown in section " Results and discussion. " Finally, conclusion and future scope are presented in section " Conclusion and future scope. "

Related work

Sentiment analysis is an important and interesting field of research in natural language processing (NLP). In recent years, many deep learning techniques have been developed for effective and efficient prediction of sentiments. In this section, we discuss the recent work accomplished in the field of sentiment analysis.

One of the most important breakthroughs in the NLP is the word embeddings [ 4 ]. In 2013, Google released a tool called “Word2Vec” to calculate word vectors. It helped to relate the similar words with each other and grab the semantic information from a sentence effectively. Another similar word embedding, namely, “GloVe” vector embedding, was researched and developed by the researchers of Stanford University [ 3 ]. These embeddings helped us to overcome major flaws in one-hot encoding, i.e., sparse distribution of words and increased dimensionality of the dataset.

The deep learning techniques have emerged in recent years, and showed excellent results in this field. Most commonly used models in NLP are CNNs [ 5 ], RNN [ 6 ] and long short-term memory [ 7 ]. The usage of word embeddings has made it possible to use CNN for texts and enables the CNN to extract local information in the sentence [ 8 ]. A two-layer CNN model was proposed by Attardi and Santos to classify the text sentences on the basis of extracted features [ 9 ], whereas Yin and Schütze used a multi-channel CNN network by forming combinations of words [ 10 ]. The detailed review on the application of CNN is presented in [ 32 ], wherein challenges in sentiment analysis are highlighted for the textual, visual and multimodal data.

LSTM and CNN–LSTM are other types of models which have been widely applied for the task of sentiment analysis. A tree-structured LSTM model was introduced by Tai et al . [ 11 ] which achieved good results in sentiment classification, whereas Wang et al. introduced a tree-structured CNN–LSTM model for dimensional sentiment analysis [ 12 ] which captured both local and long distance dependencies between words in sentences. Other than this, Wang et al. [ 13 ] also proposed an attention-based LSTM network that focuses on various parts of the sentences. Further, attention model is improved for sentiment analysis in [ 30 ]. The cognition grounded data is utilized for training the model, wherein contextual information from sentence and document level is extracted at the time of model training. The accuracy of 66.80% is achieved on Yelp14 dataset by the proposed hybrid model.

A CNN–LSTM-based model was also proposed by Zhou et al . [ 14 ] which first utilized the CNN layer to extract the sentence local features and later applied the LSTM layer in place of pooling layer to obtain the desired classification results. Sun et al . [ 15 ] applied CNN–LSTM model and hybrid deep learning algorithms to classify the Tibetan blogs which achieved the good results.

In recent years, there has been serious adaptation of attention mechanism in text analysis-based models. Yang et al . [ 18 ] combined the bidirectional RNN (BRNN) and the attention layer for the text-level classification task, whereas Long et al . [ 31 ] proposed an improved CNN and multilayered attention mechanism for the classification task of sentiment analysis. In [ 33 ], a systematic survey for the textual, visual and multimodal data is discussed for the RNN and variants of RNN such as LSTM and gated recurrent unit (GRU). The challenges and issues pertaining to sequence modeling in different modalities are highlighted.

On the basis of aforementioned literature review, we identified the potential of CNN, LSTM and the attention mechanisms for the task of sentiment analysis as per the past research. In this paper, we propose a multi-channel CNN-BiLSTM model with an attention mechanism to extract more relevant and important features to predict the sentiments from the given input. Table 1 presents the summary of the related work and highlights the benefits and drawbacks of the existing systems.

Methodology

Before passing the dataset through defined model, the data needs to be preprocessed, cleaned and represented in vector form for the model to train or test. Other than that, the data also needs to be split and tokenized. Figure  1 shows an overall dataflow of the proposed system. Firstly, dataset preprocessing and splitting is performed. Word embedding and tokenization is applied on the dataset. Training is performed using proposed classification model, and finally, accuracy and losses are visualized. Further, Fig.  2 presents the preprocessing of raw input text for the task of sentiment analysis.

Overall system model

Preprocessing of raw input text

Word embedding

To better encode the semantic and co-occurrence information, word embeddings are utilized. A word embedding is an n -dimensional dense vector of floating values and these values are generated on the basis of cosine similarity between these words. Here, n represents the vector dimensions of each embedded word. Hence, overcome the problem of sparse matrix and increased dimensions in case of one-hot encoding. There are mainly two pretrained word embeddings—(1) Word2Vec and (2) GloVe. Here, we have used the Glove embeddings of 300 dimensions having 42 billion tokens with vocab size of 1.9 million. Glove vector model obtains meaning from word-to-word co-occurrence statistical information.

Convolutional neural network

It was the Yoon Kim [ 17 ] who first used the CNN for sentence classification. It has the ability to capture semantic local information from a sentence. It can do that by using the different numbers and sizes of filters or kernels. These filters are applied to the input to extract the feature maps. Filters with different sizes are able to extract variety of information from a sentence. We usually refer to it as multi-channel CNN, i.e., we define a model with different input channels for processing n -grams of text, where n is the number of words a kernel reads at a time. Here, we have defined the model for five channels to process 3-g, 4-g, 5-g, 6-g and 8-g with equal padding. Usually, after this a pooling layer is utilized to consolidate the results from convolutional layer. Later, these pooled layers are concatenated, flattened and passed through a dense layer to classify the text. A general CNN architecture is shown in Fig.  3 [ 19 ].

Bidirectional LSTM

This paper implements the multi-channel CNN method except that in place of using pooling layer, all of the convoluted outputs are passed into two bidirectional LSTM layer to obtain the contextual information from each of the extracted local data. A long short-term memory model has the great ability to remember important information about the input as compared to its predecessor RNN, which will for the longer period of time. It overcomes the problem of vanishing gradient in RNN by using the gating mechanism. LSTM uses the input gate, output gate, forget gate and memory cell to remember the contextual information for longer period of time. These gates are composed of sigmoid layer which decides how much each component of information should go through a gate. The value of 1 means, let everything go and the value of 0 means, to block everything. A LSTM architecture is presented in Fig.  4 [ 20 ].

LSTM Cell architecture [ 20 ]

Later, it was observed that LSTM layers also tend to lose information as the sentence gets long and they only allow the information to propagate forward, i.e., the current state T output depended only on the states before T. Therefore, to extract more contextual information from a sentence, a BiLSTM model was proposed. It was inspired from the bidirectional RNN model which was proposed in 1997 by Schuster and Paliwal [ 21 ]. It not only considers the previous states but also the future states that will come after state T , which undoubtedly gives the excellent results. Figure  5 represents the BiLSTM layer model [ 22 ].

BiLSTM layer

Here, we implemented the BiLSTM layer, after the convoluted output generated from CNN and process this data to extract contextual information from the input values. Further, BiLSTM output is passed through an attention layer to generate the word scores.

Attention mechanism

It is certain that not every word of the sentence contributes equally to the meaning of a sentence. Some words have higher importance and others do not. The attention mechanism checks for such words and assigns a score to each word on the basis of their importance.

Here, we have implemented Bahdanau [ 23 ] attention. All the outputs from the BiLSTM layer are passed through this layer to assign score to each output. Bahdanau attention takes all the hidden states of last layer at once and multiply them with attention layer’s hidden state weights to assign an attention score to each input hidden state by taking a softmax of output values. After multiplying each input hidden state with attention score, a weighted sum is taken for all input hidden states. This weighted sum is called context vector and it is later passed onto the next layer. After receiving the weighted sum or adjusted hidden states, then we passed them through a single dense layer.

After performing the aforementioned operations for each channel of CNN (here in this case 5 channels), all results of each dense layer are concatenated and passed through a dense layer network to classify text either, positive or negative. The input and output dimensions for each layer for the proposed model are shown in Table 2 ; moreover, the proposed deep learning model architecture for the task of sentiment analysis is described in Fig.  6 . Here, five channels of CNNs with different kernel size and filter of 128, followed by the BiLSTM layers to extract the temporal features from text, are identified. Next, attention mechanism is applied to generate the score for the important words and all these vectors are supplied to dense layer, and further classification is performed on the final data.

Proposed deep learning model for sentiment analysis

Experimental setup and dataset

To build and train the proposed Google Colab is utilized with access to free TPU’s. In addition, Tensorflow 2.0 + and keras software libraries are used to prepare the model. To fine-tune the proposed model, various hyperparameters values are adopted to achieve the better performance. The model is trained for 20 epochs. To avoid overfitting, dropout of 0.2 is applied at second last dense layer and early-stopping is implemented with patience of three epochs. Adam optimizer with learning rate of 0.00005 and clipnorm with value “1” are used to handle the problem of exploding gradient. For loss calculation, binary cross-entropy function is applied. In case of CNN, to obtain rich collection of information, we used kernels of five different sizes—3, 4, 5, 6 and 8, wherein number of each kernel are 128. The summary of hyperparameters and their values are presented in Table 3 .

Dataset description

For this work, we have selected IMDB movie review dataset [ 16 ]. It consists of around 50 K movie reviews in which 25 k are positive labeled and remaining 25 k are negative labeled. The dataset is divided into a train/valid split ratio of 80–20. For validation, 20% of validation data is taken.

Data preprocessing

Data preprocessing is an important step before the development of any deep learning model. Before passing text into neural networks, it is necessary to clean the dataset and remove any noise in the dataset, thus to improve the performance of developed model. The IMDB movie review dataset consists sentences with URL’s, hashtags, slang words, abbreviations, emojis, stopwords, etc. The dataset persists with lots of noise. Therefore, to remove such noises such as punctuations, URL’s, hashtags and stopwords are removed from the dataset. In addition, all uppercase letters are converted into lowercase; all emojis are converted into text. Other than that, lemmatization is performed on the dataset to normalize the words. Lemmatization uses the context in which the word is being used and replaces the word with the word of similar context. Later, the data is tokenized using keras_tokenizer and padded with the sequence length of 300.

Results and discussion

We compared the proposed multi-channel CNN + BiLSTM + attention mechanism model with multi-channel CNN, LSTM, multi-channel CNN and machine learning algorithms, and found that the proposed model gives better results than the state-of-the-arts. Our model achieved a highest accuracy of 94.13% with early-stopping with patience of three epochs to prevent overfitting. Table 4 shows the performance metric in the form of accuracy based on CNN, LSTM, CNN + LSTM and proposed model on the IMDB dataset. From Table 4 , we depict that the proposed model outperforms the other models. It is also seen from Table 5 that proposed sentiment classification model’s accuracy is higher than the existing models in [ 24 , 25 , 26 , 27 , 28 , 29 ] and [ 33 , 34 , 35 , 36 ]. The increased accuracy is reported by the proposed model, i.e., more accuracy of 3.46% from [ 33 ], 5.13% from [ 34 ], 1.13% from [ 35 ] and 6.01% from [ 36 ], respectively.

Accuracy and loss of the proposed models are shown in Figs.  7 and 8 , respectively. The validation accuracy is improved after 10th epochs and reaches to the maximum of 94.31%; moreover, the validation loss is reduced to approximately 0.1.

Training and validation accuracy of proposed model

Training and validation loss of proposed model

Conclusion and future scope

In this paper, we present a multi-channel CNN with BiLSTM and attention layer. The proposed model captures both local and contextual information from the sentence. The effective and efficient features are extracted from the proposed model wherein the attention layer significantly improves the accuracy of classification model as compared with the state-of-the-arts. The proposed model achieved an accuracy of 94.13% on the IMDB dataset which shows the superiority among the other models developed in the past. The limitation of proposed model would be challenged in training using attention module which needs more data and computing resources to learn attention weights. In future, we will focus on more advanced models in CNNs and RNNs to develop the generalized model to predict the more accurate sentiments from the different datasets available in public domain.

Data availability

All data generated or analyzed during this study are included in this published article.

Zhao J, Liu K, Xu L (2016) Sentiment analysis: mining opinions, sentiments, and emotions

Pang B, Lee L, Vaithyanathan S (2002) Thumbs up? Sentiment classification using machine learning techniques. arXiv preprint cs/0205070

Pennington J, Socher R, Manning CD (2014) Glove: global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp 1532–1543

Mikolov T, Chen K, Corrado G, Dean J (2013) Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781

Fukushima K (1980) neocognitron: a self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biol Cybern 36(4):193–202

Article   MATH   Google Scholar  

Sherstinsky A (2018) Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. CoRR abs/1808.03314 (2018). arXiv preprint arXiv:1808.03314

Tang D, Qin B, Feng X, Liu T (2015) Effective LSTMs for target-dependent sentiment classification. arXiv preprint arXiv:1512.01100

Pang B, Lee L (2005) Seeing stars: exploiting class relationships for sentiment categorization with respect to rating scales. arXiv preprint cs/0506075

Attardi G, Sartiano D (2016) UniPI at SemEval-2016 Task 4: convolutional neural networks for sentiment classification. In: Proceedings of the 10th international workshop on semantic evaluation (SemEval-2016), pp 220–224

Yin W, Schütze H, Xiang B, Zhou B (2016) Abcnn: Attention-based convolutional neural network for modeling sentence pairs. Trans Assoc Comput Linguist 4:259–272

Article   Google Scholar  

Tai KS, Socher R, Manning CD (2015) Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075

Wang J, Yu LC, Lai KR, Zhang X (2019) Tree-structured regional CNN-LSTM model for dimensional sentiment analysis. IEEE/ACM Trans Audio Speech Lang Process 28:581–591

Wang Y, Huang M, Zhu X, Zhao L (2016) Attention-based LSTM for aspect-level sentiment classification. In: Proceedings of the 2016 conference on empirical methods in natural language processing, pp 606–615

Zhou C, Sun C, Liu Z, Lau F (2015) A C-LSTM neural network for text classification. arXiv preprint arXiv:1511.08630

Sun B, Tian F, Liang L (2018) Tibetan micro-blog sentiment analysis based on mixed deep learning. In: 2018 international conference on audio, language and image processing (ICALIP). IEEE, pp 109–112

Maas A, Daly RE, Pham PT, Huang D, Ng AY, Potts C (2011) Learning word vectors for sentiment analysis. In: Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pp 142–150

Kim Y (2014) Convolutional neural networks for sentence classification proceedings of the 2014 conference on empirical methods in natural language processing, EMNLP 2014, october 25–29, 2014, doha, qatar, a meeting of sigdat, a special interest group of the acl. Association for Computational Linguistics, Doha, Qatar

Yang Z, Yang D, Dyer C, He X, Smola A, Hovy E (2016) Hierarchical attention networks for document classification. In: Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies, pp 1480–1489

Gurucharan M (2020) Basic CNN architecture: explaining 5 layers of convolutional neural network. https://www.upgrad.com/blog/basic-cnn-architecture

Greff K, Srivastava RK, Koutník J, Steunebrink BR, Schmidhuber J (2016) LSTM: a search space odyssey. IEEE Trans Neural Netw Learn Syst 28(10):2222–2232

Article   MathSciNet   Google Scholar  

Schuster M, Paliwal KK (1997) Bidirectional recurrent neural networks. IEEE Trans Signal Process 45(11):2673–2681

Tavakoli N (2019) Modeling genome data using bidirectional LSTM. In: 2019 IEEE 43rd annual computer software and applications conference (COMPSAC), vol 2, pp 183–188. IEEE

Bahdanau D, Cho K, Bengio Y (2014) Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473

Bui V, Le NT, Nguyen VH, Kim J, Jang YM (2021) Multi-behavior with bottleneck features LSTM for load forecasting in building energy management system. Electronics 10(9):1026

Rehman AU, Malik AK, Raza B, Ali W (2019) A hybrid CNN-LSTM model for improving accuracy of movie reviews sentiment analysis. Multimed Tools Appl 78:26597–26613

Qaisar SM (2020) Sentiment analysis of IMDb movie reviews using long short-term memory. In: 2020 2nd international conference on computer and information sciences (ICCIS). IEEE, pp 1–4

Dong Y, Fu Y, Wang L, Chen Y, Dong Y, Li J (2020) A sentiment analysis method of capsule network based on BiLSTM. IEEE Access 8:37014–37020

Nafis NSM, Awang S (2021) An enhanced hybrid feature selection technique using term frequency-inverse document frequency and support vector machine-recursive feature elimination for sentiment classification. IEEE Access 9:52177–52192

Al Bataineh A, Kaur D (2021) Immunocomputing-based approach for optimizing the topologies of LSTM networks. IEEE Access 9:78993–79004

Long Y, Xiang R, Lu Q, Huang CR, Li M (2019) Improving attention model based on cognition grounded data for sentiment analysis. IEEE Trans Affect Comput 12(4):900–912

Diwan T, Tembhurne JV (2022) Sentiment analysis: a convolutional neural networks perspective. Multimed Tools Appl, pp 1–25

Tembhurne JV, Diwan T (2021) Sentiment analysis in textual, visual and multimodal inputs using recurrent neural networks. Multimed Tools Appl 80:6871–6910

Domadula PSSV, Sayyaparaju SS (2023) Sentiment analysis of IMDB movie reviews: a comparative study of Lexicon based approach and BERT Neural Network model. BS Thesis, May 2023

Sabba S, Chekired N, Katab H, Chekkai N, Chalbi M (2022). Sentiment analysis for IMDb reviews using deep learning classifier. In 2022 7th international conference on image and signal processing and their applications (ISPA). IEEE, pp 1–6

Atandoh P, Zhang F, Adu-Gyamfi D, Atandoh PH, Nuhoho RE (2023) Integrated deep learning paradigm for document-based sentiment analysis. J King Saud Univ Comput Inf Sci 35(7):101578

Google Scholar  

Basarslan MS, Kayaalp F (2022) Sentiment analysis with various deep learning models on movie reviews. In 2022 international conference on artificial intelligence of things (ICAIoT). IEEE, pp 1–5

Hussein SA, Bayoumi AERS, Soliman AM (2023) Automated detection of human mental disorder. J Electric Syst Inf Technol 10(1):1–10

Khairy M, Al-Makhlasawy RM (2022) A reliable image compression algorithm based on block luminance adopting deep learning for video surveillance application. J Electric Syst Inf Technol 9(1):21

Vanga J, Ranimekhala DP, Jonnala S, Jamalapuram J, Gutta B, Gampa SR, Alluri A (2023) Fault classification of three phase induction motors using Bi-LSTM networks. J Electric Syst Inf Technol 10(1):1–15

Hassanein AM, Mohamed AG, Abdullah MA (2023) Classifying blinking and winking EOG signals using statistical analysis and LSTM algorithm. J Electric Syst Inf Technol 10(1):44

IMDB Dataset. https://developer.imdb.com/non-commercial-datasets/ . Accessed 13 May 2023

Download references

No funding was received for conducting this study.

Author information

Authors and affiliations.

Department of Computer Science and Engineering, Sardar Patel University, Balaghat, MP, India

Adepu Rajesh & Tryambak Hiwarkar

Department of Computer Science and Engineering, Guru Nanak Institute of Technology, Ibrahimpatnam, Hyderabad, Telangana, India

Adepu Rajesh

You can also search for this author in PubMed   Google Scholar

Contributions

Adepu Rajesh performed the design and implementation of proposed work and prepared the initial draft of manuscript. Tryambak Hiwarkar conceptualized the work, examined and suggested the changes in the manuscript. All authors reviewed and finalized the manuscript.

Corresponding author

Correspondence to Adepu Rajesh .

Ethics declarations

Competing interests.

The authors have no competing interests to declare that are relevant to the content of this article.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Rajesh, A., Hiwarkar, T. Sentiment analysis from textual data using multiple channels deep learning models. Journal of Electrical Systems and Inf Technol 10 , 56 (2023). https://doi.org/10.1186/s43067-023-00125-x

Download citation

Received : 15 August 2023

Accepted : 10 November 2023

Published : 23 November 2023

DOI : https://doi.org/10.1186/s43067-023-00125-x

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Sentiment analysis
• Deep models

Social Media Text Sentiment Analysis: Exploration Of Machine Learning Methods

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Review Article
• Open access
• Published: 09 May 2024

The application of machine learning techniques in posttraumatic stress disorder: a systematic review and meta-analysis

• Jing Wang   ORCID: orcid.org/0009-0000-5913-2655 1 , 2   na1 ,
• Hui Ouyang 1 , 2   na1 ,
• Runda Jiao 3   na1 ,
• Suhui Cheng 1 , 2   na1 ,
• Haiyan Zhang 4 ,
• Zhilei Shang 1 , 2 ,
• Yanpu Jia 1 , 2 ,
• Wenjie Yan 1 , 2 ,
• Lili Wu 1 , 2 &
• Weizhi Liu   ORCID: orcid.org/0000-0001-6836-5522 1 , 2  

npj Digital Medicine volume  7 , Article number:  121 ( 2024 ) Cite this article

113 Accesses

1 Altmetric

Metrics details

• Human behaviour
• Population screening

Posttraumatic stress disorder (PTSD) recently becomes one of the most important mental health concerns. However, no previous study has comprehensively reviewed the application of big data and machine learning (ML) techniques in PTSD. We found 873 studies meet the inclusion criteria and a total of 31 of those in a sample of 210,001 were included in quantitative analysis. ML algorithms were able to discriminate PTSD with an overall accuracy of 0.89. Pooled estimates of classification accuracy from multi-dimensional data (0.96) are higher than single data types (0.86 to 0.90). ML techniques can effectively classify PTSD and models using multi-dimensional data perform better than those using single data types. While selecting optimal combinations of data types and ML algorithms to be clinically applied at the individual level still remains a big challenge, these findings provide insights into the classification, identification, diagnosis and treatment of PTSD.

Similar content being viewed by others

AI-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial

Causal machine learning for predicting treatment outcomes

Major depressive disorder: hypothesis, mechanism, prevention and treatment

Introduction.

Posttraumatic stress disorder (PTSD) denotes a psychological disorder characterized by delayed onset and prolonged duration, resulting from exposure to or witnessing an exceptionally threatening or catastrophic traumatic incident. The primary symptoms of PTSD as the Diagnostic and Statistical Manual of Mental Disorders 5th edition (DSM-5) noted encompass intrusive experience, persistent avoidance of stimuli, negative alterations in cognitions and mood, as well as marked alterations in arousal and reactivity related to the traumatic events 1 which can persist for extensive periods, spanning months or even years 2 . Over 70% adults encounter at least one traumatic event at some point in their lives 3 . PTSD is estimated to affect approximately 5% to 10% of the population 4 . The lifetime prevalence of PTSD varies from 1.3 to 12.2%, and the 12-month prevalence is 0.2 to 3.8% according to socio-cultural factors 5 .

Various studies have underscored the enduring detrimental impact of PTSD on an individual’s physical, mental, and social well-being, including social isolation, chronic pain, inflammation, cardiometabolic disorders, and an increased risk of chronic dementia, emphasizing the importance of identifying and predicting PTSD populations 4 , 6 . However, the psychopathology of PTSD involves a wide array of genetic, endocrine, demographic, and environmental factors that are not uniformly present in all individuals with PTSD, implying that more effective interventions may need to be tailored to specific groups or individuals.

In order to achieve effective interventions at the individual level, the most suitable analytical method for an individual’s unique bio-psycho-social characteristics is machine learning (ML). The concept of ML was originally coined by Arthur Samuel to signify the process of enabling computers to learn autonomously without explicit programming 7 . Traditional PTSD research approach is the top-down approach: formulating the hypothesis, designing the experiment, collecting the experimental data and finally deciding to accept or reject the hypothesis. ML methods can, on the one hand, obtain large amounts of data at a relatively low cost, and on the other hand, generalize alternative hypotheses through analyzing these data. That is, from multi-dimensional data such as text, scales, brain images, behavioral and physiological indicators, hidden information can be discovered, common features can be extracted, and the complex relationship between PTSD and different variables can be revealed. Thus, ML methods enable the bottom-up approach in PTSD research 8 .

ML involves a variety of algorithms, common in the field of psychology: supervised machine learning (SML), unsupervised machine learning (UML), deep learning (DL), and natural language processing (NLP). There are other algorithms such as semi-supervised learning (SSL) and reinforcement learning (RL) 9 . In PTSD, ML techniques are mainly applied in pre-diagnosis screening, identifying PTSD and its subtypes, distinguishing PTSD from other psychiatric disorders, predicting PTSD and its development trajectory, and optimizing the evaluation factors of the above four items. Previous studies have demonstrated the feasibility of using ML techniques with various data types to identify and predict PTSD, e.g. using text data to classify PTSD individuals 10 , using scales to regulate their emotion 11 and select treatment for them 12 , using biomedical data to predict early risk 13 and identify metabolomic-proteomic signatures associated with PTSD 14 . A classification summary of ML models commonly used in PTSD is shown in Fig. 1 . In recent years, neural network architectures have been used more frequently in psychiatric studies for model-level fusion, which are highly efficient in handling high-dimensional features 15 . However, due to the low interpretability of neural network algorithms, many studies choose to use more explainable ML models, such as random forest (RF) and extreme gradient boosting (XGB), to analyze data and better reveal the contribution rate of variables 16 .

The names and abbreviations of commonly used machine learning models in posttraumatic stress disorder are presented in this mind map.

In this article, we aim to review the existing literature on the use of ML techniques in the assessment of subjects with PTSD to distinguish individuals with PTSD from other psychiatric disorder or from trauma-exposed and healthy controls or to optimize the predictors of PTSD. Furthermore, we evaluated what accuracy can ML techniques achieve in the classification of people with PTSD by analyzing different types of data and provided a quality measurement of these studies.

The initial search yielded a total of 873 unique records. After title and abstract review, 766 records were excluded. With 1 report not available for full-text, 106 publications were assessed for eligibility. After full-text review, 75 records were excluded. 7 studies investigated PTSD but did not apply an ML algorithm 17 . 14 studies used ML methods but did not apply in the field of PTSD 18 . 52 studies applied a ML algorithm to differentiate PTSD subjects from controls but did not report accuracy metrics 19 and 2 reports 20 are reviews. A total of 31 studies ( n  = 210,001) were included in both quantitative and qualitative synthesis (Fig. 2 ).

Note: Flow diagram according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA2020). n  = number of studies/records/reports.

These 31 studies clearly reported what data types the ML models used and their accuracy performance metrics. 17 studies used neuroimage data 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 ; 6 studies used scale data 38 , 39 , 40 , 41 , 42 , 43 ; 3 studies used text data 10 , 44 , 45 ; 2 studies used biomedical data 13 , 46 ; and 3 studies used multi-dimensional data 47 , 48 , 49 .

Study characteristics

The characteristics of 31 included studies are shown in Table 1 . We noted an upward trend for publications in the past 4 years. 25 studies (80.6%) were published since 2018 and only 6 studies were published between 2015 and 2018 (19.4%). Most studies (74.2%; 23 of 31 articles) were conducted in the USA, China, and Canada, with 10, 7, and 6 articles, respectively. Bangladesh, Brazil, Iran, Israel, Netherlands, the Republic of Korea, Saudi Arabia and UK each contributed 1 study. A detailed comparison of regional differences in the number of publications is depicted in Supplementary Fig. 1 .

Emerging patterns in the utilization of ML techniques

Among the including studies, those employed SML algorithms follow a consistent approach that is training the ML models in a labeled dataset, iteratively assessing, contrasting, and selecting variables that can effectively discriminate between PTSD and non-PTSD cases, in order to achieve optimal accuracy on an unlabeled test dataset. In contrast, UML models are trained in an unlabeled dataset to cluster individuals and ascertain pertinent latent factors. A pair of studies employed an unsupervised learning approach to construct latent profiles utilizing data from electronic medical records 44 or self-narratives alongside psychiatric diagnoses 45 and then constructed a comparative analysis of patient characteristics among the derived profiles to identify distinctive features for classification purposes. Illustrations of frequently employed linear algorithms in the literature under investigation encompass linear kernel-based support vector machines (SVM) 21 , extreme gradient boosting (XGBoost) 13 , 40 , logistic regression (LR) 42 , elastic net (EN) 38 and least absolute shrinkage and selection operator (LASSO) regularized generalized linear model (GLM) 46 . Illustrations of frequently employed non-linear algorithms include radial basis function kernel-based SVM (RBF-SVM) 21 , 46 , alternating or hierarchical multi-label decision trees (DT) 38 , 43 , 45 , and multi-layer perceptron artificial neural networks (MLP-ANN) 23 , 28 , 44 . In some cases, supervised learning (SL) algorithms were integrated with unsupervised dimension reduction and clustering algorithms to effectively extract and discern significant features from the data. Some examples of these techniques include autoencoder and representation learning, e.g. bag of words (BoW) 44 , n-gram 45 , and recursive cluster elimination support vector machines (RCE-SVM) 32 , 34 , 35 .

Meta-analysis of classification accuracy proportions and assessment of publication bias

All of the 31 studies ( n  = 210,001) were included in the calculation of pooled estimates of classification accuracy. The overall accuracy of classification models devised by ML algorithms was 0.89 (95% confidence interval (CI) of [0.88, 0.91]) (Fig. 3 ). We identified 17 studies using neuroimage data to classify PTSD patients (pooled estimates [95%CI] = 0.86 [0.82, 0.90]), 6 studies using scales data (pooled estimates [95%CI] = 0.90 [0.84, 0.96]), 3 studies using text data (pooled estimates [95%CI] = 0.87 [0.73, 1.02]) and 2 using biomedical data (pooled estimates [95%CI] = 0.88 [0.68, 1.07]). Pooled estimate of ML models’ classification accuracy of 3 studies using multi-dimensional data is 0.96 ([95%CI] = [0.93, 1.00]) which is the highest among those obtained from other single data types. The estimates of accuracies are significantly different among subgroups ( p  ≤ 0.01) (Fig. 4 ). All the Q-values with p less than 0.05 in Fig. 3 and Fig. 4 suggested a significant heterogeneity across studies both in and among subgroups. Galbraith plot also revealed several dots outside the 95%CI lines indicating heterogeneity between studies (Supplementary Fig. 2 ). The sensitivity analysis carried out leave one out analysis by omitting each study demonstrating that no individual study affected the robust results from meta-analysis (Supplementary Fig. 3 ). Evidences of publication bias were the visually asymmetric funnel plot and the result of Egger’s test ( p  ≤ 0.01). After trim and fill adding 12 studies, the overall pooled classification accuracy proportion increased from 0.89 (95% CI [0.88, 0.91]) to 0.95 (95% CI [0.93, 0.97]) (Fig. 5 ).

Rate means the effect size in the meta-analysis, which is accuracy. 95% CI means 95% credible interval. Weight of a study is calculated by weight = 1/(se 2 + t 2 ). In this formula, “se” is the standard error of its accuracy, and “ t ” is a variable responding the level of the study based on the Q test and its effect size (i.e., accuracy).

Rate means the effect size in the meta-analysis, which is accuracy. 95% CI means 95% credible interval. Weight of a study is calculated by weight = 1/(se 2 + t 2 ). In this formula, “se” is the standard error of its accuracy, and “ t ” is a variable responding the level of the study based on the Q test and its effect size (i.e., accuracy). The pooled estimate of each subgroup is shown right below the studies in that group.

Note: The visually asymmetric funnel plot before trim and fill ( a ) were the evidence of publication bias. 12 studies were added after trim and fill ( b ). The gray funnel line represents the 95% confidence interval. The red line represents the estimated effect size. The dark blue dots represent the included studies’ effect sizes. The magenta dots represent the imputed studies’ effect sizes.

Quality assessment

As shown in Supplementary Table 2 , we evaluated the methodological features of the included studies regarding the important points to be considered in machine-learning-based studies 50 . According to the assessment tool proposed by L.F. Ramos-Lima et al. 51 (Supplementary Table 1 ), approximately 41.9% (13 out of 31 articles) of the articles included in this review employed a statistical sample or a larger sample that accurately represented the target population of the study. A majority of the studies (51.6%; 16 out of 31 articles) implemented methodologies to mitigate the influence of confounding variables, such as age, gender, or trauma type. In 64.5% of the studies (20 out of 31 articles), a rigorous and unbiased assessment conducted by independent evaluators was employed to quantify the presence of PTSD symptoms. The remaining studies utilized self-report measures, medical records, or non-blinded interviews to gather data. Additionally, a vast majority of the studies (96.8%; 30 out of 31 articles) provided comprehensive details regarding the ML techniques employed, while all of the studies (100.0%; 31 out of 31 articles) explicitly specified performance metrics such as accuracy, area under the curve (AUC), sensitivity, and specificity. A subset of the studies (25.8%; 8 out of 31 articles) presented details regarding the presence of missing data and the approach employed to address them, predominantly through imputation methods. This aspect is significant in the context of ML implementation and holds the potential to impact the outcomes based on the chosen technique. A substantial majority of the studies (93.5%; 29 out of 31 articles) employed an independent and distinct dataset to assess the performance of the generated model. Furthermore, a notable proportion of the studies (45.2%; 14 out of 31 articles) provided an account of their strategies in managing the class imbalance issue, which holds significance in calibrating and comprehending accuracy metrics. It is worth noting that a significant number of these studies maintained an equivalent sample size for both cases and controls. A comprehensive description of techniques employed for feature selection and/or hyperparameter optimization was provided in the majority of the studies (93.5%; 29 out of 31). These techniques are essential for reducing dimensionality and fostering the development of more resilient and streamlined models.

This systematic meta-analysis evaluated 31 studies used ML techniques to assess PTSD involving totally 210,001 participants in the calculation and yielded an overall estimate classification accuracy of PTSD of 89.4%. The majority of the studies (80.6%, 25 of 31 articles) applying these new methods in PTSD have been published in recent five years, indicating that the related computer technology has been increasingly mature and accessible.

The forest plot (Fig. 3 ), subgroup analysis (Fig. 4 ) and Galbraith plot (Supplementary Fig. 2 ) revealed the heterogeneity between both studies and subgroups and accuracies varied significantly. Heterogeneity between subgroups may stems from differences in data types, resulting in different methods of data collection, different levels of difficulty in sample selection, and different applicable machine learning algorithms. Additionally, different PTSD assessment tools, demographic characteristics of samples, and diverse study design are all possible sources of heterogeneity between studies. In our review, the sample size of included studies varied wildly ranging from 28 to 154,118. The effect size in the meta-analysis of this study is accuracy. Although the sample size is of concern, more attention is paid to the ratio of positive ones to the total. It determines both the effect size and the weight of studies in single proportion meta-analysis. Therefore, changes in sample size alone have little effect on the result. Its robustness was also confirmed by the sensitivity analysis (Supplementary Fig. 3 ). But small sample size sacrifices ML model performance which can be a source of heterogeneity.

The funnel plot (Fig. 5 ) was clearly asymmetrical, suggesting publication bias in the studies. This may be due to the fact that researchers tend to use models that perform well and choose to report the optimal model and its accuracy metrics.

The result of quality assessment is also noteworthy. The number of studies in the two subgroups: neuroimage and scales was relatively large, and the consequence revealed some commonalities. The process of collecting neuroimage data is time-consuming, laborious, and costly. As a result, the sample size is generally small and underrepresented. However, the collection process is carried out by professionals, which allows for better control of confounding variables and is less prone to missing data. The scale data collection is relatively simple, and a large sample size can be obtained easily. But the collection process is difficult to control by the experimenter, the confounding variables are not easy to control, and there may be a significant difference in the number of samples between PTSD and non-PTSD. Multi-dimensional data is closer to the real situation, but the problems in the process of collecting various single types of data are also concentrated, resulting in higher integrated and more complex ML algorithms and better performance.

Our study is one of the few existing systematic reviews of using ML techniques in trauma-related disorders and to conduct a meta-analysis on this topic. It quantitatively proves the effectiveness of ML technology in the field of PTSD, and provides AI-empowered (AI, artificial intelligence) evidence and ideas for the screening, diagnosis, treatment, and prognosis of PTSD. At the same time, our study learns from the published systematic reviews in this field 51 , reveals the limitations of the evidence through quantitative analysis, points out the advantages and disadvantages of ML models when using various types of data, and set some lights on the direction for better application in PTSD in the future. DL is the current trend of ML development, and it is rarely used in existing PTSD research. Its excellent processing ability for high-latitude data will play a prominent role in future PTSD research 52 .

In recent years, AI has been increasingly applied and contributing in clinical practice. Will high-performance ML models threaten the position of psychiatrists 53 ? Never. Human experts play an important role in AI+mental health. First of all, as the title suggests, ML is a technique, and the psychiatrist is the diagnostic. In Chinese, ML is “Shu” and psychiatrists are “Tao”. Their relationship is the same as method and theory. Psychiatrists make diagnoses based on the data provided by ML models who are not qualified to make. Second, ML cannot fully guarantee the quality of information. Since ML models are mostly networked, the information they collect is uneven, and human experts need to check it, so that ML models can become a reliable source of mental health information. Thirdly, there are legal risks in using ML models. It may cause patients’ privacy leakage on the one hand, and ethical issues such as misdiagnosis on the other hand without the quality to bear the corresponding legal responsibility. So human experts need to make legal guarantees 54 . Fourth, the ML model provides judgment based on the existing program, and can only be learned and judged according to the original program if the parameters are not changed. In the face of sudden new situations, the ML model does not have the ability to adjust in time, but will lower the original performance because of the emergence of this special sample. At this point, human experts can play a role in correcting deviations, thus helping the model learn new samples.

Pooled estimates of classification accuracy from multi-dimensional data (0.96 [0.93, 1.00]) are the highest comparing to those obtained from other single data types (0.86 [0.82, 0.90] to 0.90 [0.84, 0.96]). Similar results have been found in related fields such as the diagnosis and treatment of other mental disorders. For example, Lee Yena et al. had found that predictive ML models using multiple data types reached highest overall classification accuracy of 0.93 in predicting the therapeutic outcomes of depression while the models with lower-dimension data reached the proportion of 0.68-0.85 55 . Katharina Schultebraucks et al. utilized a neural network approach to prove that the integration of multi-dimensional data provides a stronger prediction of both PTSD and major depressive disorder (MDD) than one source independently with an area under curve (AUC) of 0.90 and 0.86 respectively 52 , 56 . However, the standardization in ML models reports needs to be considered so as to provide some inspirations into the consistency and comparability of the results. First, what are the architecture of the ML models? Are they single or ensembled? Supervised or unsupervised? The models in comparison should be of the same type. Second, the performance metrics of the model should be reported for different needs. For example, sensitivity should be reported for the ability to screen positive patients, specificity should be reported for the ability to screen healthy people, and the Youden index should be reported for the optimal cut-off of a questionnaire. Third, the performance evaluation methods of the model should be flexible. While accuracy and AUC are common indicators to evaluate a model, in practical application, the generalization, complexity, operability, time cost, hardware cost and other factors of the model should be considered.

ML is an effective method applied to PTSD recognition and prediction. This data-driven approach can reduce the gap between experienced and inexperienced clinicians and therefore might eliminate a proportion of the reliance on experienced clinicians in the assessment of PTSD. But there’s still a long way to go from the lab to the hospital. First, most of the existing studies use symptomatic scales of PTSD, such as PCL-5 57 , as the criterion to identify PTSD. However, the scale of subjective reporting is not accurate, and there are phenomena such as recall bias, concealment, and fatigue effect, which can lead to misreporting, omission, and false positive. Second, PTSD has a small clinical diagnosis and is often comorbid with other psychiatric disorders 58 , such as depression, anxiety, and substance abuse, resulting in a smaller positive sample available for scientific research. Third, there are gender differences in the prevalence of PTSD 59 . The available data suggest that PTSD occurs twice in women than men 4 . This means that the same is true in the data that can be learned by the ML model. As a result, the model learns the characteristics of female patients more thoroughly, the better it diagnoses possible female patients. For men, the reverse is true. The less data can be learned from male patients, the less likely it is to be diagnosed in the future. This may lead to more severe gender disparities. Fourth, influenced by cultural differences, Easterners are more introverted and Westerners are more extroverted. Existing data may lead ML models to provide similar results of gender differences. The same can happen with the localization of symptomatic scales and public health emergencies 60 . As a result, the diagnosis of PTSD has become a process that varies from time to time and from place to place and potential limitations occur in the overall prevalence. Although the DSM-5 is now the globally accepted standard, the diagnosis of PTSD cannot be generalized when it comes to the specific individual and his unique life experience in a hospital office. Therefore, ML models still need to learn a lot of knowledge, such as gender differences, cultural differences, comorbidities, public health emergencies, etc.

And our findings also implicate that the data collected from web-based methods (such as smartphone apps 42 or Global Positioning System (GPS) data derived passively from a smartphone 46 ) can be used for PTSD which brings great convenience to clinical practice. In the actual collection of data, it is inevitable to use these efficient collection methods, which will involve personal privacy, social ethics, and potential selection bias 61 . How to improve the rigor of the procedure, the protection of private information, the avoidance of social discriminatory bias, and the iatrogenic self-fulfilling prediction will be the problems that need to be solved gradually in the future 62 .

All of the included studies used SML algorithms and only two of them used unsupervised methods. SL exhibits robust capability in understanding data characteristics when provided with labeled data, enabling accurate predictions and classifications. Conversely, while UL finds numerous applications, its learning efficacy might be impacted in situations where data labeling is insufficient. Therefore, SML algorithms are prevalently favored and frequently employed for the diagnosis and detection of various disorders. Currently, for reviews both about other mental disorders and in non-psychological fields, this is the case in most of them. In a study about therapeutic outcomes in depression, twenty-four out of the total twenty-six included articles (92%) reported the utilization of SML methods 55 . In a systematic review regarding trauma-related disorders, only three out of the forty-nine included studies used UML techniques and the others (94%) all used SML techniques 51 . A review focusing on the effectiveness of ML techniques on voice disorders only considered SML algorithms because they are more commonly used in diagnosing and screening disorders 63 . In an article about artificial intelligence in identifying left ventricular scar using cardiac magnetic resonance imaging, thirty-one of the total thirty-five studies (89%) used supervised methods and significantly outperformed in sensitivity and specificity than unsupervised models 64 . UML models are commonly used for clustering and dimensionality reduction. The advantages are that the data does not need to be labeled, which avoids the difficulty of establishing PTSD diagnostic standards and effectively saves costs. Moreover, UML can be applied for anomaly detection and pattern recognition, helping to better discover the hidden characteristics, patterns and anomalies in the population. UML is also one of the paradigms that DL has been continuously exploring and innovating in recent years. Deep neural networks (DNN) use structures and methods like autoencoders represent and generate high-dimensional and unstructured data effectively (such as images, text, speech, etc.). However, the UML methods have the disadvantage of poor interpretability, and it is difficult to explain and convincing when used in the diagnosis of PTSD. At the same time, due to the lack of labels and objective functions, it is difficult for models to evaluate and adjust parameters 65 .

We conducted a meta-analytic and systematic review following the reporting checklist in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Statement (PRISMA2020) 66 (Supplementary Table 3 ). The protocol for this review was pre-registered at PROSPERO (CRD42023342042) 67 . We assess existing research endeavors that have employed ML algorithms to discern biological and phenomenological attributes with the potential to ascertain, diagnose, and prognosticate PTSD. Within the context of the meta-analysis, we compute pooled estimations of classification accuracy for ML-derived models, i.e., the proportion of correctly (vs. incorrectly) classified PTSD cases by different types of data. All analyses were based on previous published studies and thus no ethical approval and patient consent are required.

Search strategy

We searched Pubmed, Embase, Scopus, PsycINFO and Cochrane Library for publications in English. Publications in Chinese were searched through China National Knowledge Infrastructure Database (CNKI), Wanfang database, and China Science and Technology Journal Database (CSTJ) from 2000 to May 2022. We used the union set of entry terms and Medical Subject Headings (MeSH) terms as text words to search in titles and abstracts and every text word was searched as a whole. The search terms for machine learning were: ((“Machine Learning”) OR (“Deep Learning”) OR (“Hierarchical Learning”) OR (“Support Vector Machine”) OR (“SVM”) OR (“Reinforcement Learning”) OR (“Natural Language Processing”) OR (“Semi-supervised Learning”) OR (“Gaussian process”) OR (“Cross-validation”) OR (“Cross Validation”) OR (“Regularized Logistic”) OR (“Linear Discriminant Analysis”) OR (“LDA”) OR (“Random Forest”) OR (“Naïve Bayes”) OR (“Naive Bayes”) OR (“Bayesian”) OR (“Least Absolute Shrinkage And Selection Operator”) OR (“LASSO”) OR (“Elastic net”) OR (“RVM”) OR (“Relevance Vector Machine”) OR (“Pattern Recognition”) OR (“Pattern Classification”) OR (“Computational Intelligence*”) OR (“Machine Intelligence”) OR (“Knowledge Representation*”) OR (“Big Data”) OR (“Artificial Intelligence”)) AND ((“PTSD”) OR (“Post-traumatic Stress Disorder*”) OR (“Posttraumatic Stress Disorder*”) OR (“Post Traumatic Stress Disorder*”) OR (“Stress Disorder*, Post Traumatic”) OR (“Stress Disorder*, Posttraumatic”) OR (“Stress Disorder*, Post-traumatic”)). See Supplementary Data 1 for detailed information on the search terms. A comprehensive search for relevant literature will also be conducted, which will entail tracing the references of the included studies to identify potential eligible sources.

Inclusion and exclusion criteria

This review encompassed all peer-reviewed publications that investigate the impact of ML methodologies on clinical practice for PTSD. All identified publications underwent a rigorous screening process, wherein the titles and abstracts were assessed for eligibility. The inclusion criteria were deliberately broad in order to enhance the search sensitivity. The specific inclusion criteria are summarized as follows: (1) PTSD was the dependent variable in the study. (2) Data related to the use of ML methods for PTSD are reported. (3) There is a clear type of data processing such as text, neuroimage, scale, biomedicine and multi-dimensional types. (4) Accuracy (ACC) metrics of ML algorithms are reported.

Studies were excluded if one of the following conditions is met: (1) The study is one of the following types: qualitative studies, editorials, letters, case studies, comments, notes, reviews, protocols or meta-analyses. (2) There is an explicit reference to non-PTSD related brain injury and genetic and biological studies. (3) No ML methods are used to study PTSD. (4) PTSD is not the dependent variable in the study. (5) The study does not report ML model’s accuracy metrics. (6) The study is not in Chinese or English.

In the meta-analytic quantitative synthesis, only peer-reviewed publications meeting the following criteria were included: reporting a measure of classification accuracy and providing information on the sample size of the analysis set. Non-refereed publications were excluded from the meta-analytic synthesis.

Data extraction

The titles and abstracts of the identified articles were independently screened by two researchers (JW and HO). Subsequently, they acquired and thoroughly examined the full texts of potentially relevant articles. In cases of disagreement, RJ acted as the final arbiter for decision-making. Throughout the primary and secondary screening processes, all activities were overseen by WY and LW to ensure quality control. Data extracted from the articles encompass four key domains: (1) study characteristics (authors, year of publication, data type, diagnostic PTSD tool); (2) participant information (count type, sample size, diagnosis); (3) models (ML model, accuracy, other measures). (4) Additional information deemed pertinent to the study, including event details and accompanying commentary have also been extracted. Requests for supplementary information were made to the authors of two studies, with one author promptly responding and providing the requested data. SC and HZ assisted in the interpretation of the findings. Subsequently, all authors engaged in comprehensive discussions regarding the results and made significant contributions to the final version of the manuscript.

In cases where multiple measures of classification performance metrics were reported in the studies, all measures were initially extracted. Subsequently, the accuracies were either extracted directly or calculated, and these values were utilized in the quantitative analyses. If a study reported results from multiple ML models (e.g., using different ML methods, sets of features), the classification accuracy of the best performing model used in the hold-out set or the balanced accuracy was included in the calculation of pooled estimates.

Quality analysis

We conducted a quality assessment using the tool which Luis Francisco Ramos-Lima proposed in 2020 51 , as there was no such instrument in ML studies before that. For studies using ML techniques in healthcare, the specific dimensions of quality analysis are described in Supplementary Table 1 . It is divided into nine domains relating to the sample quality and effect size of a study. The first three dimensions are sample representativeness, confounding variables, and outcome assessments which are the most concerned aspects of individuals in healthcare researches implement with ML techniques. The remaining six dimensions can evaluate the quality and affect the performance of a ML approach: the model description, its accuracy or other performance metrics, how missing data and class imbalance problems had been handled, whether the model had been tested on unseen data, and whether the results were optimized using hyperparameter optimization or feature selection procedures. This review does not include studies that are of low quality or clearly have design and method flaws. The results are represented in the Results Section and shown in Supplementary Table 2 . ZS and YJ independently conducted quality assessments of the risk of bias for the studies included in this review. In the event of discrepancies between the two reviewers, WL consulted to resolve any disagreements.

Data synthesis and meta-analytic procedures

All analyses were based on previous published studies and thus no ethical approval and patient consent are required. We extracted both qualitative and quantitative data as mentioned above from each selected study. Then, we conducted meta-analysis of studies where data availability allows summary estimation for accuracy of 95% confidence intervals. For each included research, the precise count of accurately classified occurrences was ascertained by multiplying the documented proportion of classification accuracy by the sample size (n) and rounding the resulting product to the nearest integer value. Forest plots were used for the presentation of results. A random effects model was applied to estimate the overall accuracy of all the included studies. All statistics were calculated in STATA MP 17 with 0.05 as the significance level.

We chose the Q-value as an indicator of heterogeneity, with p less than 0.05 suggesting a significant heterogeneity across studies. Funnel plots and Egger’s test 68 were adopted to detect publication bias for there were more than 10 studies reporting the primary outcomes 69 .

The studies were categorized into subgroups based on the processed data type, namely text, neuroimage, scale, biomedical, and multidimensional data. We used the STATA MP 17 to conduct subgroup analysis that reported the accuracies vary among subgroups. Forest plots were again used to display the results. We also carried out a sensitivity analysis by excluding one study each time to explore whether the results were driven by a study with an extreme result.

Data availability

Data collected and used in this meta-analysis can be requested from the corresponding author in response to reasonable requests.

Association, A. P. Diagnostic and Statistical Manual Of Mental Disorders: DSM-5 5th edn, 271–280 (American Psychiatric Publishing, 2013).

Mota, N. et al. Course and predictors of posttraumatic stress disorder in the canadian armed forces: a nationally representative, 16-year follow-up study. Can. J. Psychiatry 66 , 982–995 (2021).

Article   PubMed   PubMed Central   Google Scholar  

Benjet, C. et al. The epidemiology of traumatic event exposure worldwide: results from the World Mental Health Survey Consortium. Psychol. Med. 46 , 327–343 (2016).

Article   CAS   PubMed   Google Scholar  

Yehuda, R. et al. Post-traumatic stress disorder. Nat. Rev. Dis. Prim. 1 , 108–114 (2015).

Google Scholar  

Shalev, A., Liberzon, I. & Marmar, C. Post-traumatic stress disorder. N. Engl. J. Med. 376 , 2459–2469 (2017).

Article   PubMed   Google Scholar  

Song, H. et al. Association of stress-related disorders with subsequent neurodegenerative diseases. JAMA Neurol. 77 , 700–709 (2020).

Kim, Y. K. & Na, K. S. Application of machine learning classification for structural brain MRI in mood disorders: critical review from a clinical perspective. Prog. Neuro-Psychopharmacol. Biol Psychiatry 80 , 71–80 (2018).

Article   Google Scholar  

Danmin, M. et al. Fusion and model construction of multi-objective measurement of personality traits. J. Air Force Med. Univ. 43 , 131–135 (2022).

Jan, Z. et al. The role of machine learning in diagnosing bipolar disorder: scoping review. J. Med. Internet Res. 23 , e29749–e29767 (2021).

Sawalha, J. et al. Detecting presence of PTSD using sentiment analysis from text data. Front. Psychiatry 12 , 132–146 (2022).

Christ, N. M., Elhai, J. D., Forbes, C. N., Gratz, K. L. & Tull, M. T. A machine learning approach to modeling PTSD and difficulties in emotion regulation. Psychiatry Res. 297 , 113712–113720 (2021).

Held, P. et al. Who will respond to intensive PTSD treatment? A machine learning approach to predicting response prior to starting treatment. J. Psychiatr. Res. 151 , 78–85 (2022).

Schultebraucks, K. et al. Forecasting individual risk for long-term posttraumatic stress disorder in emergency medical settings using biomedical data: a machine learning multicenter cohort study. Neurobiol. Stress 14 , 100297–100306 (2021).

Kuan, P. F. et al. Metabolomics analysis of post-traumatic stress disorder symptoms in World Trade Center responders. Transl. Psychiatry 12 , 174–180 (2022).

Article   CAS   PubMed   PubMed Central   Google Scholar  

Khoo, L. S., Lim, M. K., Chong, C. Y. & McNaney, R. Machine learning for multimodal mental health detection: a systematic review of passive sensing approaches. Sens. (Basel) 24 , 348–412 (2024).

Dupont, T., Kentish-Barnes, N., Pochard, F., Duchesnay, E. & Azoulay, E. Prediction of post-traumatic stress disorder in family members of ICU patients: a machine learning approach. Intensive Care Med 50 , 114–124 (2024).

Shaw, S. A., Ward, K. P., Pillai, V. & Hinton, D. E. A group mental health randomized controlled trial for female refugees in Malaysia. Am. J. Orthopsychiatry 89 , 665–674 (2019).

Shiba, K. et al. Uncovering heterogeneous associations of disaster-related traumatic experiences with subsequent mental health problems: A machine learning approach. Psychiatry Clin. Neurosci. 76 , 97–105 (2022).

Samuelson, K. W. et al. Mental health and resilience during the coronavirus pandemic: a machine learning approach. J. Clin. Psychol. 78 , 821–846 (2022).

Schultebraucks, K. & Chang, B. P. The opportunities and challenges of machine learning in the acute care setting for precision prevention of posttraumatic stress sequelae. Exp. Neurol. 336 , 113526–113541 (2021).

Saba, T. et al. Machine learning for post-traumatic stress disorder identification utilizing resting-state functional magnetic resonance imaging. Microsc Res. Tech. 85 , 2083–2094 (2022).

Zhu, Z. et al. Combining deep learning and graph-theoretic brain features to detect posttraumatic stress disorder at the individual level. Diagnostics 11 , 11080000–11080012 (2021).

Yang, J. et al. Using deep learning to classify pediatric posttraumatic stress disorder at the individual level. BMC Psychiatry 21 , 535–544 (2021).

Sheynin, S. et al. Deep learning model of fMRI connectivity predicts PTSD symptom trajectories in recent trauma survivors. Neuroimage 238 , 118242–118254 (2021).

Li, L. et al. Hippocampal subfield alterations in pediatric patients with post-traumatic stress disorder. Soc. Cogn. Affect Neurosci. 16 , 334–344 (2021).

Chen, Z. et al. Neural connectome prospectively encodes the risk of post-traumatic stress disorder (PTSD) symptom during the COVID-19 pandemic. Neurobiol. Stress 15 , 100378–100387 (2021).

Nicholson, A. A. et al. Classifying heterogeneous presentations of PTSD via the default mode, central executive, and salience networks with machine learning. NeuroImage Clin. 27 , 102262–102276 (2020).

Lanka, P. et al. Supervised machine learning for diagnostic classification from large-scale neuroimaging datasets. Brain Imaging Behav. 14 , 2378–2416 (2020).

Kim, Y. W. et al. Riemannian classifier enhances the accuracy of machine-learning-based diagnosis of PTSD using resting EEG. Prog. Neuropsychopharmacol. Biol. Psychiatry 102 , 109960–109967 (2020).

Harricharan, S. et al. PTSD and its dissociative subtype through the lens of the insula: anterior and posterior insula resting-state functional connectivity and its predictive validity using machine learning. Psychophysiology 57 , 72–94 (2020).

Salminen, L. E. et al. Adaptive identification of cortical and subcortical imaging markers of early life stress and posttraumatic stress disorder. J. Neuroimaging 29 , 335–343 (2019).

Rangaprakash, D., Dretsch, M. N., Katz, J. S., Denney, T. S. & Deshpande, G. Dynamics of segregation and integration in directional brain networks: Illustration in soldiers with PTSD and neurotrauma. Front Neurosci. 13 , 803–825 (2019).

Nicholson, A. A. et al. Machine learning multivariate pattern analysis predicts classification of posttraumatic stress disorder and its dissociative subtype: a multimodal neuroimaging approach. Psychol. Med. 49 , 2049–2059 (2019).

Rangaprakash, D. et al. Identifying disease foci from static and dynamic effective connectivity networks: Illustration in soldiers with trauma. Hum. Brain Mapp. 39 , 264–287 (2018).

Jin, C. et al. Dynamic brain connectivity is a better predictor of PTSD than static connectivity. Hum. Brain Mapp. 38 , 4479–4496 (2017).

Liu, F. et al. Characterization of post-traumatic stress disorder using resting-state fMRI with a multi-level parametric classification approach. Brain Topogr. 28 , 221–237 (2015).

Nicholson, A. A. et al. Differential mechanisms of posterior cingulate cortex downregulation and symptom decreases in posttraumatic stress disorder and healthy individuals using real-time fMRI neurofeedback. Brain Behav. 12 , 2441–2461 (2022).

Ramos-Lima, L. F. et al. Identifying posttraumatic stress disorder staging from clinical and sociodemographic features: a proof-of-concept study using a machine learning approach. Psychiatry Res. 311 , 114489–114496 (2022).

Jiang, T. et al. Toward reduced burden in evidence-based assessment of PTSD: a machine learning study. Assessment 28 , 1971–1982 (2021).

Ge, F., Li, Y., Yuan, M., Zhang, J. & Zhang, W. Identifying predictors of probable posttraumatic stress disorder in children and adolescents with earthquake exposure: A longitudinal study using a machine learning approach. J. Affect Disord. 264 , 483–493 (2020).

Leightley, D., Williamson, V., Darby, J. & Fear, N. T. Identifying probable post-traumatic stress disorder: applying supervised machine learning to data from a UK military cohort. J. Ment. Health 28 , 34–41 (2019).

Wshah, S., Skalka, C. & Price, M. Predicting posttraumatic stress disorder risk: a machine learning approach. JMIR Ment. Health 6 , e13946–e13957 (2019).

Magoc, D., Magoc, T., Tomaka, J. & Morales-Monks, S. Automatic identification of firefighters with post-traumatic stress disorder based on demographic characteristics and self-reported alcohol consumption. JUS 10 , 198–203 (2016).

Zafari, H., Kosowan, L., Zulkernine, F. & Signer, A. Diagnosing post-traumatic stress disorder using electronic medical record data. Health Inform. J. 27 , 159–178 (2021).

He, Q., Veldkamp, B. P., Glas, C. A. & de Vries, T. Automated assessment of patients’ self-narratives for posttraumatic stress disorder screening using natural language processing and text mining. Assessment 24 , 157–172 (2016).

Lekkas, D. & Jacobson, N. C. Using artificial intelligence and longitudinal location data to differentiate persons who develop posttraumatic stress disorder following childhood trauma. Sci. Rep. 11 , 10303–10312 (2021).

Shahid, F. et al. Leveraging free-hand sketches for potential screening of PTSD. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 4 , 94:01–94:22 (2020).

Worthington, M. A., Mandavia, A. & Richardson-Vejlgaard, R. Prospective prediction of PTSD diagnosis in a nationally representative sample using machine learning. BMC Psychiatry 20 , 532–541 (2020).

Tahmasian, M. et al. Differentiation chronic post traumatic stress disorder patients from healthy subjects using objective and subjective sleep-related parameters. Neurosci. Lett. 650 , 174–179 (2017).

Passos, I. C. et al. Machine learning and big data analytics in bipolar disorder: a position paper from the International Society for Bipolar Disorders Big Data Task Force. Bipolar Disord. 21 , 582–594 (2019).

Ramos-Lima, L. F., Waikamp, V., Antonelli-Salgado, T., Passos, I. C. & Freitas, L. H. M. The use of machine learning techniques in trauma-related disorders: a systematic review. J. Psychiatr. Res 121 , 159–172 (2020).

Schultebraucks, K., Yadav, V., Shalev, A. Y., Bonanno, G. A. & Galatzer-Levy, I. R. Deep learning-based classification of posttraumatic stress disorder and depression following trauma utilizing visual and auditory markers of arousal and mood. Psychol. Mede 52 , 957–967 (2022).

Sezgin, E. Artificial intelligence in healthcare: complementing, not replacing, doctors and healthcare providers. Digit Health 9 , 1–5 (2023).

Ngiam, K. Y. & Khor, I. W. Big data and machine learning algorithms for health-care delivery. Lancet Oncol. 20 , e262–e273 (2019).

Lee, Y. et al. Applications of machine learning algorithms to predict therapeutic outcomes in depression: a meta-analysis and systematic review. J. Affect Disord. 241 , 519–532 (2018).

Schultebraucks, K. & Galatzer-Levy, I. R. Machine learning for prediction of posttraumatic stress and resilience following trauma: an overview of basic concepts and recent advances. J. Trauma Stress 32 , 215–225 (2019).

Ouyang, H. et al. The increase of PTSD in front-line health care workers during the COVID-19 pandemic and the mediating role of risk perception: a one-year follow-up study. Transl. Psychiatry 12 , 180–186 (2022).

Yan, W. et al. The impact of isolation on comorbidity of PTSD symptoms and depression: evidence from PTRP-5-6 in China. BMC Public Health 24 , 21–30 (2024).

Liu, N. et al. Prevalence and predictors of PTSS during COVID-19 outbreak in China hardest-hit areas: Gender differences matter. Psychiatry Res. 287 , 112921–112927 (2020).

Zhou, Y. et al. The prevalence of PTSS under the influence of public health emergencies in last two decades: a systematic review and meta-analysis. Clin. Psychol. Rev. 83 , 101938–101948 (2021).

Jing, W., Zhilei, S. & Weizhi, L. Application of machine learning in mental health. J. Nav. Med Univ. 44 , 1145–1153 (2023).

Kelly, C. J., Karthikesalingam, A., Suleyman, M., Corrado, G. & King, D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 17 , 195–203 (2019).

Al-Hussain, G., Shuweihdi, F., Alali, H., Househ, M. & Abd-Alrazaq, A. The effectiveness of supervised machine learning in screening and diagnosing voice disorders: systematic review and meta-analysis. J. Med Internet Res 24 , 72–90 (2022).

Jathanna, N. et al. Diagnostic utility of artificial intelligence for left ventricular scar identification using cardiac magnetic resonance imaging—a systematic review. Cardiovasc Digit Health J. 2 , S21–S29 (2021).

Eckhardt, C. M. et al. Unsupervised machine learning methods and emerging applications in healthcare. Knee Surg. Sports Traumatol. Arthrosc. 31 , 376–381 (2023).

Page, M. J. et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372 , n71-n79 (2021).

Wang, J. et al. Machine learning methods to discriminate posttraumatic stress disorder: a protocol of systematic review and meta-analysis. Digit Health 10 , 20552076241239238–20552076241239243 (2024).

Irwig, L., Macaskill, P., Berry, G. & Glasziou, P. Bias in meta-analysis detected by a simple, graphical test. Graphical test is itself biased. BMJ 316 , 470–471 (1998).

CAS   PubMed   PubMed Central   Google Scholar  

Egger, M., Smith, G. D., Schneider, M. & Minder, C. Bias in meta-analysis detected by a simple, graphical test. BMJ 315 , 629–634 (1997).

Download references

Acknowledgements

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: General Program of National Natural Science Foundation of China (32071086); Science and Technology Supply Project (2020JY17); Military Postgraduate Funding Projects (202346-295).

Author information

These authors contributed equally: Jing Wang, Hui Ouyang, Runda Jiao, Suhui Cheng.

Authors and Affiliations

Lab for Post-traumatic Stress Disorder, Faculty of Psychology and Mental Health, Naval Medical University, 200433, Shanghai, China

Jing Wang, Hui Ouyang, Suhui Cheng, Zhilei Shang, Yanpu Jia, Wenjie Yan, Lili Wu & Weizhi Liu

The Emotion & Cognition Lab, Faculty of Psychology and Mental Health, Naval Medical University, 200433, Shanghai, China

Graduate School, PLA General Hospital, 100853, Beijing, China

Department of Health Care, The First Affiliated Hospital of Naval Medical University, 200433, Shanghai, China

Haiyan Zhang

You can also search for this author in PubMed   Google Scholar

Contributions

J.W., H.O., R.J. and S.C. contributed to the writing of this article and are the co-first authors. H.Z., Z.S., Y.J., W.Y. and J.W. contributed to the article review and revised the article. W.L. and L.W. led the whole study, including putting this study forward and carrying out the study; they are the co-corresponding authors.

Corresponding authors

Correspondence to Lili Wu or Weizhi Liu .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary material, rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Wang, J., Ouyang, H., Jiao, R. et al. The application of machine learning techniques in posttraumatic stress disorder: a systematic review and meta-analysis. npj Digit. Med. 7 , 121 (2024). https://doi.org/10.1038/s41746-024-01117-5

Download citation

Received : 23 September 2023

Accepted : 23 April 2024

Published : 09 May 2024

DOI : https://doi.org/10.1038/s41746-024-01117-5

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

A Review on Public Acceptance Towards New Education Policy in India Based on Sentiment Analysis Using Machine Learning Approaches

• Conference paper
• First Online: 14 May 2024
• Cite this conference paper

• Deepali Chaudhary 13 , 14 ,
• Ritam Dutta 14 &
• Papri Ghosh 15  

Part of the book series: Lecture Notes in Networks and Systems ((LNNS,volume 916))

Included in the following conference series:

• International Conference on Information and Communication Technology for Competitive Strategies

Education plays a pivotal role in the lives of individuals, as well as in the progress of nations and the global community. As the Information and Communication Technology (ICT) continuously advances and improves with each successive generation, education becomes the bedrock upon which the younger generation builds their future. In recognition of this, Government of India has formulated a new education policy in 2020. In line with this holistic development, we have steered a comprehensive study utilizing a diverse range of Machine Learning (ML) and Deep Learning (DL) algorithms to survey public Sentiment Analysis (SA) and their acceptance towards the National Education Policy (NEP) in India. Additionally, we have undertaken a detailed analysis of recent literature survey pertaining to the NEP, illustrated the research gap and finally conducted a comparative assessment of the findings. The accuracy has been recorded maximum at 94% using LSTM model in comparison of NB, SVM, KNN models.

This is a preview of subscription content, log in via an institution to check access.

Access this chapter

• Available as PDF
• Read on any device
• Instant download
• Own it forever
• Available as EPUB and PDF
• Compact, lightweight edition
• Free shipping worldwide - see info

Tax calculation will be finalised at checkout

Purchases are for personal use only

Institutional subscriptions

Athavale VA, Bansal R (2021) A review on NEP 2020 and its influence on academics 24

Google Scholar  

Kurien A, Chandramana SB (2020) Impact of NEP 2020 on higher education

Pooja M, Bhalla R (2022) A review paper on the role of sentiment analysis in quality education

Verma H, Kumar A (2021) NEP 2020 of India: a theoretical analysis. Int J Bus Manage Res 9

Biswas R, Vyas N, Baskar M (2021) Sentiment analysis on national education policy change 2020(12):11–18

Kaurav R, Suresh K, Narula S (2020) NEP 2020: qualitative analysis and twitter mining analysis 12:4–13

Sonwal P (2021) NEP 2020 and higher education. Int J Educ, Modern Manage, Appl Sci Soc Sci 3(3):55–59

Malhotra N, Goyal T (2021) Sentiment analysis using twitter information flow about the NEP introduced in India in 2020. Int J Manag 11(12):2411–2416

Meena G, Mohbey KK, Mahrishi M (2023) A ML-based approach to analyze public sentiment on Indian new education policy 2020

Bhagat B, Dandge S (2023) A literature review on sentiment analysis using ML in education domain

Kumari S (2020) NEP 2020 challenges to teachers’ education. Int J Appl Res 6(10):420–424

Rayees A, Ahmad R, Tahseena J (2023) Changing role of teacher educators in view of NEP 2020. J Xi'an Univ Architect Tech 15(1):144–156

Jadhav N (2023) Issues and challenges of NEP 2020 implementation in teaching and education. Int J Recent Engg Res Develop 10(3):188–191

Ahangar SD, Ayub M (2022) NEP-2020 and teacher education-some issues. Int J Creative Res Thoug 10(4):440–443

Mahanta B (2023) Recommendations and challenges of teacher education in NEP-2020. J Emerging Tech Innovative Res 10(5):88–92

Kumar A (2022) Importance of national education policy-2020 in imparting education. MHRD doc 6(2):137–152

MathSciNet   Google Scholar  

Gaurav S, Narula S, Tiwari P (2021) Theoretical extension of the NEP 2020 using Twitter mining. J Content, Comm Comm 13(7):16–26

Mabunda B, Ajoodha R (2021) Sentiment analysis of student textual feedback to improve teaching. Inter Discipl Res Tech Manag 1:643–651

Lazrig I, Sean L (2022) Computer information and decision management using machine learning sentiment analysis to evaluate learning impact. Informat Syst Educat J 20(3):01–07

Qaiser S (2021) A comparison of ML techniques for sentiment analysis. Turkish J Comp Math Educ 12(3):1738–1744

Sangeetha K (2021) Sentiment analysis of student feedback using multi-head attention fusion model of word & context embedding for LSTM. J Amb Intell Humanized Comp 12(3):4117–4126

Chauhan G, Agrawal P, Meena Y (2019) Aspect-based sentiment analysis of students’ feedback to improve teaching learning process. IICT Intell Syst, 259–266

Lalata J, Gerardo B (2019) A sentiment analysis model for faculty comment evaluation using ensemble machine learning algorithms. In: Proceedings of 2019 International Conference on Big Data Engineering, 13, 68–73

Gutiérrez G, Ocha A (2018) Analyzing students reviews of teacher performance using support vector machines by a proposed model, intelligent computing systems. Comm Comp Inf Sc 820:114–124

https://www.education.gov.in/sites/upload_files/mhrd/files/NEP_Final_English_0.pdf , Last accessed 2023/09/22

https://static.pib.gov.in/WriteReadData/userfiles/file/NEPBookletFinalFO2Z.pdf , Last accessed 2023/09/20

Gill S, Vrontis D, Singh G (2022) Transformative quality in higher education institutions: conceptualization, scale development and validation. J Bus Res 138:275–285

Article   Google Scholar  

Saini M, Kaur M (2021) Analysing the tweets to examine the behavioral response of Indian citizens over the approval of NEP 2020. Int J Edu Dev 82:102356

Kumar A, James WB, Cobanoglu C, Cavusoglu M (2021) NEP 2020: a roadmap for India 2.0-Ed. Glocer 21(4):134–144

Geddam MS (2022) Ideological and philosophical transformation of India education policies and NEP 2020. Int J Multidiscip Educ Res 11(3):6–12

Aithal PS, Shubhrajyotsna A (2020) Analysis of the Indian National Education Policy 2020 towards achieving its objectives. Munich Personal RePEc Archive

Muralidharan K, Shanmugan K, Klochkov Y (2022) The NEP 2020, digitalization and quality of life in India: some reflections 12:75–81

Download references

Author information

Authors and affiliations.

Department of CSE, Banasthali Vidyapith, Jaipur, Rajasthan, 304022, India

Deepali Chaudhary

Department of CSE, Poornima University, Jaipur, Rajasthan, 303905, India

Deepali Chaudhary & Ritam Dutta

Department of CSE, Narula Institute of Technology, Kolkata, West Bengal, 700109, India

Papri Ghosh

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Ritam Dutta .

Editor information

Editors and affiliations.

Global Knowledge Research Foundation, Ahmedabad, Gujarat, India

Nottingham Trent University, Nottingham, UK

Mufti Mahmud

University of Peradeniya, Delthota, Sri Lanka

Roshan G. Ragel

Department of CSE, SNS College of Technology, Coimbatore, Tamil Nadu, India

Rights and permissions

Reprints and permissions

Copyright information

© 2024 The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.

About this paper

Cite this paper.

Chaudhary, D., Dutta, R., Ghosh, P. (2024). A Review on Public Acceptance Towards New Education Policy in India Based on Sentiment Analysis Using Machine Learning Approaches. In: Joshi, A., Mahmud, M., Ragel, R.G., Kartik, S. (eds) ICT: Cyber Security and Applications. ICTCS 2022. Lecture Notes in Networks and Systems, vol 916. Springer, Singapore. https://doi.org/10.1007/978-981-97-0744-7_38

Download citation

DOI : https://doi.org/10.1007/978-981-97-0744-7_38

Published : 14 May 2024

Publisher Name : Springer, Singapore

Print ISBN : 978-981-97-0743-0

Online ISBN : 978-981-97-0744-7

eBook Packages : Engineering Engineering (R0)

Share this paper

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research