A Novel Framework for Mental Illness Detection Leveraging TOPSIS-ModCHI-Based Feature-Driven Randomized Neural Networks

Abstract

Mental illness has emerged as a significant global health crisis, inflicting immense suffering and causing a notable decrease in productivity. Identifying mental health disorders at an early stage allows healthcare professionals to implement more targeted and impactful interventions, leading to a significant improvement in the overall well-being of the patient. Recent advances in Artificial Intelligence (AI) have opened new avenues for analyzing medical records and behavioral data of patients to assist mental health professionals in their decision-making processes. In this study performance of four Randomized Neural Networks (RandNNs) such as Board Learning System (BLS), Random Vector Functional Link Network (RVFLN), Kernelized RVFLN (KRVFLN), and Extreme Learning Machine (ELM) are explored for detecting the type of mental illness a user may have by analyzing the random text of the user posted on social media. To improve the performance of the RandNNs during handling the text documents with unbalanced class distributions, a hybrid feature selection (FS) technique named as TOPSIS-ModCHI is suggested in the preprocessing stage of the classification framework. The effectiveness of the suggested FS with all the four randomized networks is assessed over the publicly available Reddit Mental Health Dataset after experimenting on two benchmark multiclass unbalanced datasets. From the experimental results, it is inferred that detecting the mental illness using BLS with TOPSIS-ModCHI produces the highest precision value of 0.92, recall value of 0.66, f-measure value of 0.77, and Hamming loss value of 0.06 as compared to ELM, RVFLN, and KRVFLN with a minimum feature size of 900. Overall, utilizing BLS for mental health analysis can offer a promising avenue toward improved interventions and a better understanding of mental health issues, aiding in decision-making processes.

1. Introduction

The World Health Organization estimates that suicide claims the lives of over 800,000 people each year, making it the fourth most common cause of death for those aged between 15 to 19 [1]. Amidst the COVID-19 pandemic lockdown, there has been a notable increase in mental health challenges and their portrayal on social media platforms. Economic instability and isolation brought on by the pandemic have long-term impacts on people’s mental health and general well-being. This phenomenon has ramifications for clinical psychologists and academic researchers. The reports from August 2021 definitively state that, in England, there were around 1.6 million people on waiting lists for mental health services [2]. Furthermore, around eight million individuals were unable to access specialist help because they did not meet the criteria for being considered sick enough. This situation highlights the necessity of automating the detection of mental health issues from social media data, where individuals freely express their thoughts, beliefs, and emotions [3]. Users are often inclined to share their mental health difficulties or illnesses anonymously on various social media or in online health communities. These online health communities can serve as a platform for sharing empathy and establishing connections with others who are experiencing similar symptoms. On social media, users also frequently look for health information on their symptoms in order to self-diagnose [4]. These massive reservoirs of user-generated data on social media platforms present an emerging alternate for examining behavioral patterns to predict the mental health of the person. The possible effects of untreated mental distress could be lessened by these predictive capabilities, which could enable prompt interventions from healthcare providers and support systems.

1.1. Research Gaps and Motivations

Questionnaires and in-person consultations with psychological consultants are the conventional methods of detecting mental disease. Despite being a very effective approach, very few patients profit from it due to a lack of enough specialists. New opportunities have been opened by recent advancements in Artificial Intelligence (AI) and Machine Learning (ML) to assist mental health professionals to identify mentally sick people by analyzing their behavioral data available in social media [5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20]. Several works utilizing traditional ML algorithms, such as Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), K-Nearest Neighbor (KNN), Naïve Bayes (NB), and Support Vector Machine (SVM) have contributed to identifying mental disorders from data available on social media [5,6,7,8,9]. The majority of these studies concentrate on creating a predictive model for a single, particular illness, treating it as a binary classification problem. However, in the real scenario, differentiating between several mental illnesses, perhaps in a comorbidity scenario, is a more difficult challenge that necessitates for a multiclass classification solution. More recently various Deep Learning (DL) techniques including Convolutional Neural Networks (CNNs) [10,11,12,13,14] Long Short-Term Memory (LSTM) [13,15,16,17], Recurrent Neural Network (RNN) [17,18], and transformers-based DL models [19,20] have been leveraged to figure out a user’s mental condition from their shared information on social media. Despite the immense capability of the deep structures, the majority of the networks suffer from the time-consuming training procedure due to the inclusion of a large number of connecting parameters in filters and layers with several hyper-parameters and complex structures [21,22]. The intricate nature of the deep structure also poses a significant challenge for theoretical analysis, leading researchers to primarily concentrate on tweaking parameters and adding more layers using more and more powerful computing resources in pursuit of improved accuracy [21,22]. In addition, it goes through a complete retraining process if its structure is inefficient to accurately reflect the system.

In contrast to Deep Neural Networks (DNNs), Single Layer Feed Forward Neural Networks (SLFNs) have also drawn interest from researchers for addressing a variety of regression and classification problems due to their simplistic topologies and universal approximation capabilities. The conventional method of training SLFNs using a Back Propagation (BP)-based iterative approach often encounters challenges such as local minima, sensitivity to learning rate, and slow convergence [23,24]. Further, to tackle these challenges, innovative Randomized Neural Networks (RandNNs) like Extreme Learning Machine (ELM) [25] and Random Vector Functional Link Network (RVFLN) [26,27] have been introduced in the literature. These networks assign arbitrary weights and biases for the hidden layer, while the weights for the output layer are computed analytically. Due to the non-iterative training procedure, these networks not only offer fast training speed, but also provide the generalization capability in function approximation [22,26,27]. Based on the idea of RVFLN, another RandNN termed as Broad Learning System (BLS) is proposed in [21,28] as an alternative way of learning deep features. A flat network is used to set up the BLS, where the initial inputs are relocated and kept as mapped features in feature nodes. These are then changed by a nonlinear activation function to create enhancement nodes through which the BLS’s structure is expanded in a wide sense. The most notable aspect of BLS as a SLFN is its use of mapped feature nodes to build enhancement nodes, which introduces a better feature representation capacity. In contrast to the DNN, BLS does not require a gradual search for the optimal parameters of the models using BP [21]. Finding the weights of the linkages between the nodes of hidden and output layers is achieved in a single step during the learning process of BLS by conducting matrix inversion. Although RandNNs are successfully used for binary and multiclass classifications in a number of fields [22,28], they are rarely used for text classification (TC), particularly in the field of mental health analysis.

1.2. Contributions

With an initiative to bridge the research gaps discussed in the previous section, this study offers a novel solution by developing a RandNN-based text classifier that can efficiently identify a wide array of mental illness, including depression, suicide, anxiety, schizophrenia, autism, post-traumatic stress disorder (PTSD), alcoholism, bipolar disorder, and a neutral category. In contrast to several studies that concentrate on detecting the presence or absence of a single mental health condition, this model aims to classify users into a range of potential mental illnesses with the nuanced expressions and linguistic patterns found in user-generated text by utilizing four RandNNs such as BLS [21,28], RVFLN [26,27], KRVFLN [29], and a ELM [25]. To boost the performance of the RandNNs during handling the text documents with unbalanced class distributions, a hybrid FS technique named as TOPSIS-ModCHI is suggested in the preprocessing stage of the classification framework. Unlike traditional chi-square, which merely uses the highest chi-square value to tank features, ModCHI, being an improved version of it, focuses on selecting features with low repetition and high relevance to class [30,31]. It assigns a score to each feature by dividing the amount of relevant documents of each class by the total number of documents in the dataset and then selects k number of highly scored features, resulting in a better handle of multilabeled text documents with unbalanced class distributions. In [30], the authors have analyzed the performance of four ML techniques, such as linear SVM, DT, RF, and KNN with ModCHI for addressing the multiclass unbalanced TC problem. By experimenting with the benchmark Reuters TC dataset, the effectiveness of the techniques is validated by comparing with four other FS techniques. The work is extended by the author in [31] by assessing the performance of RVFLN with ModCHI for TC, including 20 Newsgroups, one more benchmark dataset. In both the previous approaches of application of ModCHI, the ideal feature size is decided by assessing the performance of each classifier with 10 different feature sizes of ModCHI and then taking the minimal feature size producing the average of those 10 assessments as the ideal one. Being motivated by the use of TOPSIS for ideal feature size selection of ReliefF-based FS in [32], this research presents a further extension of the previous two works [31,32] by utilizing TOPSIS to determine the ideal feature size k of ModCHI rather than arbitrarily determining this value. With 10 different k values considered within a range of 10 to 1500, k number of highly scored features is picked by ModCHI for each RandNN-based classifier, and its performance is evaluated using four criteria such as precision, recall, F-measure, and Hamming loss for each k value. Then TOPSIS is used to rank the feature size and to select the optimal feature size for the classifier. The effectiveness of the suggested FS with all the four RandNNs is assessed over the publicly available Reddit Mental Health Dataset after experimenting on two benchmark multiclass unbalanced datasets, such as Reuters and 20 Newsgroups. Based on the experimental findings, it can be concluded that, when compared to RVFLN, KRVFLN, and ELM with a minimum feature size of 900, BLS with TOPSIS-ModCHI yields better results for each assessment criterion when used to detect mental illness.

The objective of this research is to investigate the effectiveness of RandNNs to tackle the challenges in Mental Health Analysis for unbalanced, multiclass classification. This paper contributes to the field by presenting a novel application of RandNN with a hybrid FS technique, with supporting empirical evidence. The paper’s primary contribution can be summarized as follows:

• Developing a text classifier using RandNN for efficiently categorizing multiclass unbalanced text documents.
• Selecting an optimal set of features for the classifier by utilizing a hybrid TOPSIS-ModCHI-based FS technique.
• Assessing the performance of four RandNNs, including BLS, RVFLN, KRVFLN, and ELM, to determine whether a user exhibits linguistic markers indicative of a specific mental illness by analyzing the user’s posts on social media.
• Addressing mental illness detection as an unbalanced multiclass TC task.
• Analyzing the impact of hidden layer size and activation functions on the overall performance of each RandNN is evaluated by experimenting with various hidden layer sizes and three different activation functions.

The rest of the paper is presented as follows: Section 2 presents a summary of the literature highlighting the role of ML in the identification of mental health disorders. Section 3 provides a thorough description of the techniques used as well as the schematic layout of the proposed classification model. Section 4 presents the results obtained from the model implementation and their analysis. Finally, Section 5 concludes the work, acknowledging the limitations of the study and outlining the potential avenues for future research directions.

2. Literature Review

In recent years, an ever-growing number of people have turned to popular social media platforms like Twitter, Facebook, Reddit, and Instagram to express their thoughts and connect with others instantly. This trend has led to the generation of extensive social data, which holds valuable insights into individuals’ interests, emotions, and behavior patterns [4]. Social media has had a profound impact on the way individuals self-identify as having a mental health condition and how they connect with others who share similar experiences. People frequently turn to platforms like Reddit and Twitter to ask about treatment options and side effects, feeling less isolated and stigmatized in the process. Analyzing these social media interactions can offer valuable insights into the most pressing concerns of patients, sometimes even surpassing the insight available to their doctors [3,4]. Researchers are moving beyond just looking at how often social media is used and are now using more advanced methods to analyze usage patterns. These methods include delving into the linguistic style, emotional content, and interaction patterns within social networks to predict specific mental health conditions associated with social media posts. In recent years, there has been a surge of interest in various ML techniques within the research community interested in mental health analysis.

The authors of [5] used DT, KNN, SVM, and an ensemble technique to create a binary classification model to conduct a depression analysis on Facebook posts. DT is found to achieve the highest accuracy among the tested models across experiments. Another study has explored the performance of MLP, LR, SVM, and RF for depression analysis on Reddit posts using a Linear Discriminant Analysis (LDA)-based FS technique [6]. Following the selection of key linguistic features, the TF-IDF technique is applied to extract word frequencies. Then, LDA selects the pertinent features, which are fed as input to the binary classifiers. After experimenting with single and combined feature sets, it was found that the MLP classifier using combined feature sets outperforms the other models, achieving an accuracy of 91% and F1-score of 93%. In [7], five ML techniques, including NB, DT, RF, SVM, and KNN, are used to identify three types of mental disorders, such as anxiety, depression, and stress, from the questionnaire prepared from 349 participants. Owing to the imbalanced nature of the dataset, the models are compared based on F1-score, and NB is found to surpass other models. Another comparison of six traditional ML classifiers with three FS techniques and utilizing a SMOTE-based data balancing strategy reveals the superior performance of the AdaBoost classifier with the SelectKbest FS technique for depression detection on a sample dataset prepared by the author in [8]. The performance of NB and SVM is compared in [9] for depression detection from short social media texts using a hybrid FS technique. The proposed FS technique uses the chi-based filter approach in its first stage to rank the features and then uses the whale optimization algorithm in its second phase to retrieve the optimal subset of features. The superiority of the proposed FS technique is validated by experimenting with four benchmark short text datasets.

Recent achievements of deep learning in various NLP tasks have spurred academics to develop models for mental disorder identification utilizing several DNNs, including CNN, LSTM, RNN, and others. In [10], a four-layered CNN model has shown better performance compared to a XGBoost model to classify six types of mental disorders based on the posts of the users submitted in Reddit’s mental health communities. In this study, instead of considering it as a multiclass classification problem, six classifiers are designed for detecting six types of mental illness. Kour & Gupta in [12] have suggested a hybrid CNN-biLSTM model for predicting depressive persons by analyzing their Twitter data. Comparing the hybrid DL model to the CNN and RNN models, the experimental investigations show that the hybrid model obtains the greatest accuracy of 94.28%. Another depression detection model using LSTM is proposed in [13]. The model has employed symptom-based features derived from a one-hot process with Linear Discriminant Analysis (LDA) for feature selection in place of traditional word frequency-based features. Through testing on the Norwegian dataset, the model outperformed three traditional ML models, including LR, DT, and SVM, as well as a CNN model. In [14], a simple four-layer CNN using Word2vec vectorizer is shown to perform better at identifying mental disorders than the traditional XGBoost model. The model was evaluated using user posts available on Twitter and Reddit, suggesting its importance of examining multiple social media platforms to gain a more comprehensive view of public opinion on mental health. Six types of mental illness have been identified by the study, which has addressed it as a multiclass classification problem. When compared to SVM, KNN, LR, RF, AdaBoost, and MLP models, the effective performance of a LSTM model for stress detection is presented in [15] over a self-prepared dataset created using a questionnaire. In [16], authors have explored the use of two traditional ML models such as, KNN, RF, and a LSTM-based DL model for identifying 12 different types of mental disorders. Assessing and comparing the performance of the three models on a mental health disorder dataset collected from Kaggle, the RF model is found to achieve better outcomes compared to the other two models. In [17], another DL model using a transformer and a simple architecture has shown its effectiveness on detecting depression and suicidal ideation from social media content in comparison to traditional ML models. An optimized RNN using a dynamic stabilization technique is proposed in [18] for automatic diagnosis of mental health conditions. The time domain and statistical features extracted using the dual domain feature extraction method are given as input to the RNN model. Experiments on the OSMI dataset from different age groups show that the model classifies mental illness and healthy individuals with 98% and 99.5% accuracy, respectively, when compared to CNN and SVM models. In [19], authors have investigated the automatic classification of nine types of mental disorders from Reddit posts using transformer-based DL models such as BERT, RoBERT, along with XLNet. The NB classifier is employed to choose the most informative features from each type of mental disorder to feed into the DL models. This study has focused on identifying more mental illnesses rather than one or two specific ones, performing general text analysis and classification by posts rather than by individuals or groups. Another hybrid DL model integrating CNN with a pretrained sentence BERT is utilized for detecting depressed Reddit users by utilizing the SMHD data in [20]. By capturing the semantic meaning of Reddit posts using SBERT and then identifying their behavioral patterns using CNN, the model is able to achieve an accuracy of 86% by outperforming other ML models such as CNN, LSTM, BERT, RoBERT, XLNET, LSVM, LR, and XGBoost.

Table 1. A survey on ML and DL models used for identifying mental illness.

Reference	Author (Year)	ML Model	FS Technique	Dataset	Type of Mental Illness Addressed	Evaluation Metrics
[5]	Islam et al. (2018)	DT, KNN, SVM, ensemble	---	Facebook data	Depression	Accuracy
[6]	Tadesse et al., (2019)	MLP, LR, SVM, RF	LDA	Reddit	Depression	Accuracy, F1-score
[7]	Priya et al., (2020)	NB, DT, RF, SVM, KNN	---	DASS-21	Anxiety, depression, stress	Accuracy, precision, recall, specificity, F1-score
[8]	Zulfiker et al., (2021)	KNN, AdaBoost, Gradient boosting (GB), Extreme GB, Bagging, Weighted voting	SelectKBest, mRMR, Boruta	Self-prepared dataset	Depression	Sensitivity, specificity, precision, F1-score, AUC
[9]	Priya & Karthika, (2023)	NB, SVM	Hybrid FS using Chi and Whale Optimization	Four bench mark short text datasets such as Stack Overflow, Short Messaging Service (SMS), Sentiment Labeled Sentences (SLS), and Sentiment 140 dataset	Depression	Accuracy, F1-scores, sensitivity
[10]	Kim et al., (2020)	CNN, XGBoost	---	Reddit	Depression, anxiety, bipolar, borderline personality disorder, schizophrenia, autism	Accuracy, precision, recall, F1-score
[12]	Kour & Gupta (2022)	CNN-biLSTM, CNN, RNN	Embedding layer of CNN	Twitter data	Depression	Precision, recall, F1-score, accuracy, specificity, AUC
[13]	Uddin et al., (2022)	LSTM, LR, DT, SVM, CNN	LDA	Young user’s text data obtained from ung.no Norwegian	Depression	Precision, recall, F1-score, support
[14]	Ang & Venkatachala, (2023)	CNN, XGBoost	---	Reddit, Twitter	Depression, anxiety, bipolar, BPD, schizophrenia, autism	Precision, recall, F1-score, accuracy
[15]	Bhavani & Naveen, (2024)	LSTM, KNN, RF, SVM, LR, AdaBoost, MLP	---	Self-prepared dataset through questionaries	Stress	Accuracy, precision, recall, F1-score
[16]	Alkahtani (2024)	LSTM, KNN, RF	---	Mental disorder dataset collected from Kaggle	ADHD, ASD, loneliness, MDD, OCD, PDD, PTSD, anxiety, BD, eating disorder, psychotic depression, and sleeping disorder	Accuracy, precision, recall, F1-score
[17]	Ezerceli & Dehkharghani, (2024)	RNN, CNN, LSTM, BERT, SVM, NB, RF, LR, DT	---	Reddit	Depression/suicidal ideation	AUC, precision, recall F1-score
[18]	Revathy et al., (2024)	Dynamically Stabilized RNN, CNN, SVM	Siberian tiger optimization algorithm	OSMI dataset	Mentally ill for getting treatment or not	F1-score, accuracy
[19]	Dinu & Moldovan, (2021)	BERT, XLNet, RoBERT	NB	Reddit	Schizophrenia, bipolar, depression, anxiety, obsessive-compulsive disorders, eating disorders, autism, post-traumatic stress disorder, attention-deficit/hyperactivity disorder	Precision, Recall, F1-score
[20]	Chen et al., (2023)	SBERT-CNN, CNN, LSTM, BERT, RoBERT, XLNET, LSVM, LR, XGBoost	---	SMHD dataset including Reddit posts	Depression	Accuracy, Precision, Recall, F1-score
Our approach		BLS, ELM, RVFLN, K-RVFLN	TOPSIS-ModCHI	Reddit	Depression, suicide, anxiety, schizophrenia, autism, PSTD, alcoholism, bipolar, neutral	Pmicro, Rmicro, Fmicro, Haming loss (HL), true positive rate

presents a summary of ML and DL models used in identifying different types of mental illness via social media posts.

Even with DL models’ tremendous upside, their primary drawback is the complexity of their deep structure, which leads to a laborious training process and requires researchers to concentrate mostly on adjusting their parameters and building more layers with even more potent computational resources [21,22]. As an alternative to these DL models, researchers are turning to RandNNs for a variety of binary and multiclass classification tasks because of their non-iterative training process and comparatively simple structure [26,27,28,29]. Chauhan & Tiwari (2022) have explored four RandNNs, such as RVFLN, KRVFLN, BLS, and fuzzy BLS, in [29] to address the multilabel classification problem. The model exhibits competitive performance as measured by metrics such a HL and average precision with less training time when tested on the Bibtex dataset with 159 classes, Emotions, and Scene dataset each including six classes. Another application of RVFLN for unbalanced multiclass TC is suggested by Behera & Dash (2022) in [31]. For enhancing the performance of RVFLN, a ModCHI-based FS technique is suggested in their work. When tested on the Reuters dataset containing 90 classes and 20 Newsgroups dataset with 20 classes, the model has shown superior performance in terms of precision, recall, and HL value compared to four traditional ML techniques. An ELM supplemented with a regularization factor termed as RELM is applied for TC in [33]. Using singular value decomposition on the TF-IDF matrix, the relevant features are chosen to feed to RELM. Experimenting on two benchmark TC datasets, such as the Reuters-21578 dataset with 10 classes and the WebKB dataset with four classes, the RELM-based TC model has shown better performance compared to ELM, BP, and SVM-based models. In [34], the performance of ELM along with three traditional Ml models such as SVM, KNN, and NB are compared for TC. The authors of this paper have suggested a modified frequency-based term weighting strategy to enhance the TC task by capturing the important terms present in the imbalanced text datasets. Experiments on three benchmark TC datasets show that the suggested FS technique with SVM classifier performs better than the other three models in terms of F1-score, recall, and precision. Assessing the performance of an ensemble ELM with an information entropy-based FS approach in [35] has shown better TC performance compared to weighted ELM, RELM, ELM, and three traditional ML techniques when experimented on Reuters-52 dataset with 52 classes, 20 Newsgroup with 20 classes, and WebKB with 4 classes. In [36], it is demonstrated how using max-min document frequency-based FS has improved the performance of an ELM-based TC system in categorizing the ninety different classes of the Reuters-21578 dataset. The effectiveness of combining a transformer model like BERT with BLS is highlighted in [37] to achieve both high accuracy and computationally efficiency in text emotion classification. The model has used BERT for feature extraction and BLS for efficient and fast learning, leading to improved emotion classification. As a convincing substitute for DL models, Du et al. (2020) have demonstrated the benefits of BLS for accurate and computationally efficient TC in [38]. Both word significance and sequence information is learnt by the suggested BLS for TC. When evaluated on 13 diverse datasets covering sentiment analysis and TC, the proposed model showed a considerable improvement in accuracy and training efficiency compared to the LSTM model. A node slicing BLS-based TC framework is suggested in [39]. An enhancement layer reflecting global information is developed based on the feature layer, once it is created by including rich words through one to one correspondence between the words and feature node groups. After that, a few nodes from feature layer groups are turned on and combined with the enhancement layer to produce outputs using ridge regression. The suggested BLS has shown better performance compared to CNN-BLS, LSTM-BLS, and ELM models when tested for sentiment analysis over Reddit sentiment and Twitter sentiment, each one containing three classes. A review on different types of RandNN used for TC, the type of FS technique used in those models, and the type of dataset specifying the number of classes used for validation of those models is shown in

Table 2. A review on RandNNs used for text classification.

Reference	Author (Year)	Type of RandNN	FS Technique	Dataset (Number of Classes)
[29]	Chauhan & Tiwari (2022)	RVFLN, KRVFLN, BLS, Fuzzy BLS	---	Bibtex (159), Emotions (6), Scene (6)
[31]	Behera & Dash (2022)	RVFLN, LSVM, RF, DT, MLKNN	ModCHI, Chi, MI, Tf-idf	Reuters-21578 (90) 20 Newsgroups (20)
[33]	Zheng et al. (2013)	RELM, ELM, BP,SVM	Singular value decomposition applied on TFIDF matrix	Reuters-21578 (10), WebKB (4)
[34]	Sabbah et al. (2017)	SVM, KNN,ELM, NB	Modified frequency-based term weighting scheme	Reuters-21578 (8) 20 Newsgroups (20), WebKB (7)
[35]	Li et al., (2018)	AE1WELM, Weighted ELM, RELM, ELM, SVM, KNN, NB	Information entropy	Reuters-52 (52) 20 Newsgroups (20), WebKB (4)
[36]	Behera & Dash (2021)	ELM	Max-min document frequency	Reuters-21578 (90)
[38]	Du et al., (2020)	Recurrent BLS, Gated BLS, LSTM	---	New Years resolutions (10), Political-media (9), Twitter sentiment (6), Objective-sentence (5), Apple-Twitter sentiment (4), Corporate messaging (4), Progressive tweet (4), Weather sentiment (4), Claritin-October twitter (3), Airline-sentiment (3), Tweet-global warming (3), Electronic sentiment (2), Books sentiment (2)
[39]	Liu et al., (2022)	NSBLS, ELM, CNN-BLS, LSTM-BLS	---	Reddit sentiment (3), Twitter sentiment (3)
Our approach		BLS, RVFLN, KRVFLN, ELM	TOPSIS-ModCHI	Reuters-21578 (90) 20 Newsgroups (20), Reddit Mental Health (9)

.

Based on the foregoing discussion, it is concluded that the bulk of ML models suggested for mental health analysis have focused primarily on a specific illness, addressing it as a binary classification task. This study attempted to get over this constraint by approaching mental illness detection as an unbalanced multiclass TC task. Despite the advantages of RandNNs over DL models as delivered by several researchers, the review has also revealed their infrequent application in mental health analysis. Again, less work is done related to the efficient FS of RandNNs for TC. Therefore, we proposed a TC framework using RandNN for efficiently identifying different categories of mental illness and a hybrid TOPSIS-ModCHI-based FS technique to select the optimal set of features for the classifier.

3. Materials and Methods

The fundamental concept of RandNN, with the detailed working principle of the suggested TOPSIS-ModCHI-based RandNN classification model, is covered in this section.

3.1. Randomized Neural Networks

RandNNs were put forward as a counterpart to gradient descent-based NNs to overcome their shortcomings of a highly time consumable learning process, local minima, sensitivity to learning rate, and so on. These networks have a naturally random character since they use a non-iterative process to determine the output layer’s weight and maintain all other weights constant during training [23]. Because of this randomness, RandNNs can learn faster with fewer adjustable settings and without requiring high end hardware [23,24]. ELM, RVFLN, and BLS are some popular RandNNs that have been successfully used for solving various classification and prediction problems [21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39]. Figure 1 presents the basic structural differences among ELM, RVFLN, and BLS.

3.1.1. Extreme Learning Machine (ELM)

The Extreme Learning Machine (ELM) algorithm, developed by Huang [25], is a powerful SLFN that is widely utilized for tasks such as classification and regression. It uses single step training with randomly initialized weights and biases between the input and hidden layers. Only the output weights are analytically computed using the Moore-Penrose generalized inverse of the hidden layer’s output matrix. This approach distinguishes ELM from conventional NNs, which frequently use an iterative training process, by allowing it to learn from training data in a single step. ELM has been used in various text classification problems [33,34,35,36] due to its rapid training, minimal need for parameter tuning intervention, and its ability to support a variety of activation functions.

3.1.2. Random Vector Functional Link Neural Network (RVFLN)

Like ELM, RVFLN is another RandNN with three layers representing the input, output, and hidden layer. After being received by the input layer, the input data are processed by the neurons of the hidden layer using randomly assigned weights. Finally, the output layer produces the result of the computation. The only difference in the architecture of ELM and RVFLN is that the output layer receives the original features as input through direct linkages. The direct links help in improving the generalization ability of the NN. The network also has the advantage of not using the iterative training steps [22,26]. The details of RVFLN-based text classification are detailed in [31].

3.1.3. Board Learning System

As an extension to RVFLN, another RandNN, termed as BLS, is proposed in [21] as a substitute for learning deep features. In contrast to RVFLN, the BLS maps the input layer feature matrix to feature layer F, and the feature layer F is mapped to the enhancement layer E by nonlinear activation functions. The output layer is connected to both the feature layer and the enhancement layer. BLS aims to provide an alternating approach to deep learning. Unlike DNN, BLS does not require a gradual search for the optimal parameters of the model using BP. In order to determine the weights of the linkages between the hidden and output layer, the BLS learning process only involves the one step matrix inversion.

The BLS network possesses a unique horizontal expansion and vertical fix, setting it apart from conventional deep neural networks. With a significantly reduced number of layers, its distinctive structure is illustrated in Figure 2. If X is the input feature matrix represented as [X₁, X₂, X₃, …. X_n]. The feature node F_i in feature layer maps the feature matrix from the input layer using the following equation:

$$F_i={\mathrm{\theta }}_i\left(X{W}_i+B_i\right), i=1, 2, \dots .n$$

(1)

where X act as feature matrix, and W_i and B_i act as weight and bias from the input layer to the feature layer. Here, ${\mathrm{\theta }}_i$ acts as nonlinear activation function. F_i acts as feature node in feature layer F. All feature nodes can be represented as F_n = {F₁, F₂, F₃ …. F_n}. The feature nodes act as input to the enhancement layer E, producing the enhancement nodes using the following equation:

$$E_j={\alpha }_j\left(F_n{W}_j+B_j\right), j=1, 2, \dots .m$$

(2)

where F_n act as feature node from the feature layer. W_j and B_j act as weight and bias used between the feature layer and enhancement layer. Here, ${\alpha }_j$ acts as a nonlinear activation function. All enhancement nodes can be represented as E_m = [E₁, E₂, E₃, … E_m]. The output node in output layer is obtained using the following:

$$Y=\left[ {(F}_1, F_2, F_3, \dots \dots F_n) \right| {(E}_1, E_2, E_3, \dots \dots E_m) ] W_{n+m}$$

(3)

Here, $W_{n+m}$ represents the total weights used in between feature layer to output layer and in between enhancement layer to output layer.

If $\left[{(F}_1,F_2,F_3,\dots \dots F_n)\right|{(E}_1,E_2,E_3,\dots \dots E_m\left)\right]$ is represented as $[H_1,H_2,H_3\dots \dots H_m]$, then the output node Y can be further calculated as follows:

$$Y=\left[ F_n\right|E_m ] W_{n+m}$$

(4)

If A = $\left[ F_n\right|E_m ]$, then

$$Y=A{W}_{n+m}$$

(5)

In BLS, the parameters such as W_i, W_j, B_i, and B_j are randomly assigned. As the values of matrix A and Y are already known in the learning process, so the weights represented by $W_{n+m}$ are analytically computed using the Moore–Penrose generalized inverse as follows:

$$W_{n+m}=A^{+}Y$$

(6)

Here, A⁺ represents the pseudo inverse of matrix A.

3.2. Proposed RandNN-Based Text Classification Model Utilizing TOPSIS-ModCHI-Based FS

This section provides a comprehensive guide on the steps involved in utilizing RandNNs for text classification with the TOPSIS-ModCHI-based FS method. Figure 3 depicts the proposed RandNN-based text classification model.

As the main objective of this study is to analyze the performance of RandNN for detecting mental disorder of a person from his/her social media posts, a publicly available dataset named “Reddit Mental Health Dataset” is used for the validation of the model. As this problem refers to a multiclass unbalanced text classification problem, so to enhance the performance of the classifier, TOPSIS-ModCHI based, a hybrid FS technique is used in the classification framework.

Data preprocessing is a crucial step in text classification that involves transforming raw data into a format that can be used for text classification. In this process, each document in the corpus goes through the steps of converting all the text to lower case or upper case, eliminating unnecessary characters, abbreviations, emojis, and punctuation. Additionally, stop words and duplicate entries are removed, and spelling errors are corrected. Lemmatization and tokenization further refine the data by converting words to their base forms and splitting text into individual units. Stemming is a linguistic process that involves breaking down a word into its root form by removing any prefixes, infixes, or suffixes. This technique is commonly used to identify the core meaning of a word and to group together words with similar roots. To carry out this process, the most widely recognized and commonly used stemmer, known as the Porter stemmer, has been employed in this study. The Porter stemmer algorithm is a well-established technique that effectively removes suffixes from words while preserving the core meaning. Lemmatization is a linguistic process that is used to determine the base form or root word of a given word. It performs a thorough analysis of the morphology of a word and provides the root form of all its inflectional forms. This approach is more accurate than stemming, which may not always produce the correct root word.

Once the data pre-processing is complete, the raw text data are transformed into feature vectors through a process of feature extraction. This vital step involves the removal of any redundant features, which helps to decrease the dimensionality of the attributes. The process of converting raw text into a feature vector involves the use of the Term Frequency-Inverse Document Frequency (TF-IDF) method. This method not only helps to identify the frequency of individual terms in a document but also the importance of these terms about the entire corpus. By assigning higher weights to words that are rare in the corpus but frequent in the document, the TF-IDF method helps to create a feature vector that captures the most relevant information and can be used for further analysis.

Following the feature extraction, a hybrid FS technique is utilized in the next step to include the pertinent set of features by narrowing the feature space. The TF-IDF matrix calculated in the previous step and the total count of required k features are given as input to this FS technique. For each class of each attribute, the chi-square value is computed using the following equation:

$$Chi-square\left(t_i,c_j\right)={\displaystyle \frac{N{\left(PS-RQ\right)}^2}{\left(P+R\right)\left(Q+S\right)\left(P+Q\right)\left(R+S\right)}}.$$

(7)

where N presents the number of documents in a corpus, P specifies the documents in class c_j containing terms t_i, Q is the documents in other class with term t_i, R is the documents of class c_j not containing terms t_i, and S is the documents in the other class without term t_i. Then, for each feature in the feature set, each feature’s class number with highest chi value and the maximum chi value for each feature are found. Following it, a triplet containing feature list, highest chi value, and class number is formed for each class. Based on their chi value, the triplets are arranged in descending order. The number of document belonging to same class is multiplied by the number of features, and then the result is divided by the total number of document in the corpus. This yields a score based on each class’s feature and the quantity of documents in that class. After that, the feature list of those class numbers is chosen if the value shown in the triplet corresponds to the score that was determined. This approach will increase the number of documents with fewer features in the training set while giving significant attention to relevant classes. Another challenge involved with this technique is identifying its optimal k value. Instead of randomly setting this value, further, the ideal feature size of ModCHI is decided by TOPSIS, a Multi Criteria Decision Making (MCDM) approach, as specified in [32]. The performance of the classifier is observed with different k value of ModCHI with respect to different criteria to form the decision matrix, which is then passed as input to TOPSIS. The highest ranked k value obtained by applying the steps of TOPSIS is considered as the ideal one. The flow chart of TOPSIS-based ranking is presented in Figure 4.

Then the structure of four RandNNs, such as ELM, RVFLN, KRVFLN, and BLS, is established by specifying the number of nodes used in their different layers and the activation functions used for those nodes. KRVFLN is an extension of RVFLN that uses the kernel function for mapping input features to the hidden layers, so as to increase the stability of standard RVFLN [29]. Kernel functions are a powerful tool for measuring the similarity between two vectors, providing a greater value for similar vectors compared to dissimilar ones. When using the polynomial kernel, two vectors will yield the maximum value if they are parallel, even if they are not similar in length. In this study, we have used the polynomial kernel as follows:

$$K\left(X,Y\right)={\left(X^T Y+C\right)}^d$$

(8)

where X and Y represent the input vectors and C is the constant term.

For each RandNN, according to the feature size k, that number of nodes is fixed in the input layer, and the output layer size is fixed to the total classes. For ELM, RVFLN, and KRVFLN, hidden layer size is set to the summation of one-third of the feature size with the total number of output nodes [31]. With k number of selected features and y number of output classes, the hidden layer size m is calculated as follows:

$$m=\left({\displaystyle \frac{k}{3}}\right)+y .$$

(9)

The same size is also taken for the feature and enhancement layer in BLS. During the training of ELM, RVFLN, and KRVFLN, the weights between the input and hidden layers are initialized randomly, and the hidden layer output matrix is calculated using a suitable activation function. Then, the output layer weights are calculated in a single step by finding the pseudo inverse of the hidden layer output matrix. Similarly in case of BLS, the weights used in between input and feature layer and in between feature and enhancement layer are initialized randomly, and then the output layer weights are calculated in a single step using Equation (6). Finally the performance of the model is accessed using Hamming loss and the micro averaged value of precision, recall, and F-measure as calculated as follows:

$$HL={\displaystyle \frac{1}{NL}}\sum _{i=1}^N\sum _{j=1}^{\mathrm{C}}XOR\left({target}_{ij},{predicted}_{ij}\right).$$

(10)

where N is the number of samples, C is the number of classes, target_ij is the true class label for instance i and class j, and predicted_ij is the predicted label for instance i and class j.

$$P_{micro}={\displaystyle \frac{{\sum }_{i=1}^C{tp}_i}{{\sum }_{i=1}^C{tp}_i+{fp}_i}}.$$

(11)

where tp is the true positive, and fp is the false positive.

$$R_{micro}={\displaystyle \frac{{\sum }_{i=1}^C{tp}_i}{{\sum }_{i=1}^C{tp}_i+{fn}_i}}$$

(12)

$$F_{micro}={\displaystyle \frac{2\ast p_{micro}\ast r_{micro}}{p_{micro}+r_{micro}}}$$

(13)

4. Results

This section presents a comprehensive comparison of four RandNNs leveraging ModCHI-based FS for classifying mental health conditions from social media text by experimenting with the publicly available “Reddit Mental Health Dataset”. Reddit is a social media platform comprising user-generated content organized into themed communities called “subreddits” [40]. Users can post, share, and discuss content ranging from news and entertainment to niche interests. The original dataset contains posts from 28 subreddits from the year 2018 to 2020. This paper has used a small piece of data including 75,558 posts for analysis from the original dataset that comprises of 8 different classes of mental illnesses, which are depression, suicide, anxiety, schizophrenia, autism, PTSD, alcoholism, and bipolar disorder, and a set of neutral data class of the year 2019 only. Before experimenting over the mental health dataset, the performance of the proposed model is analyzed over two benchmark datasets used for text classification, such as Reuters-21578 and 20 Newsgroups. The trials are run on a Core i3 computer with 4GB of RAM running Ubuntu 16.04.6 LTS. Python 3.8.18 is used to put the algorithms and techniques into implementation.

Table 3. Description of the dataset.

Dataset Name	Total Number of Documents	Total Number of Class	Total Number of Training Documents	Total Number of Testing Documents
Reddit Mental Health	75,559	9	60,446	15,112
Reuters-21578	10,788	90	7769	3019
20 Newsgroups	5226	20	4180	1046

shows the number of classes, along with the total number of documents of each dataset. The distribution of documents in the Reddit Mental Health Dataset based on their respective classes is shown in

Table 4. Class-wise distribution of documents in the Reddit Mental Health Dataset.

Class Name	Number of Documents	Number of Training Documents	Number of Testing Documents
Depression	33,554	26,838	6716
Suicide	14,592	11,679	2913
Anxiety	13,213	10,590	2623
Schizophrenia	1550	1224	326
Autism	1401	1111	290
PSTD	1397	1127	270
Alcoholism	942	755	187
Bipolar	910	702	208
Neutral	7999	6420	1579
Total	75,558	60,446	15,112

, and its graphical representation using a pie chart is shown in Figure 5. A sample of a Reddit posts with their corresponding class/subreddit is shown in

Table 5. Sample user’s post related to each class from the Reddit Mental Health Dataset.

Sl.no	Post	Subreddit
1	HELP If life was a dream, what would we have to do to wake up from it.	Depression
2	How should I end my life Girls get a ton of attention so they either ignore me or treat me like shit. The human race sucks. Im going to be alone forever	Suicide
3	I got a job interview on Monday at 11. Not going as I’m too anxious Thats it. First interview after studying and getting my diploma. I’m too nervous that ill screw it over. Advice needed	Anxiety
4	I put out a lit cigarette on my leg. My thoughts are uncontrollable. The violent urges and thoughts are uncontrollable now, they just keep coming.	Schizophrenia
5	Is autism genetic? I’m a teenager with autism spectrum disorder. My dad also has it. If I eventually have kids, what are the chances that they will also be autistic?	Autism
6	I am afraid to attribute to my trauma every bad thing that I feel. Hey guys, Just wanted to know if anyone can relate to. It has been roughly a year since my traumatic episode took place and ever since my life has became a mess.	PSTD
7	Been an alcoholic for past 3 years Like my title says I’ve been struggling with drinking for past few years. My drink of choice is vodka. I occasionally drink beer as well.	Alcoholism
8	Enjoy life with this one weird trick. Yes, you are bipolar--or maybe you are the SO of a bipolar person. Here is one trick that will make 2019 better than any other year has been.	Bipolar
9	I’m meeting up with one of my besties tonight! Can’t wait!!	Neutral

. Referring to the hold-out validation approach used in several works, particularly in the context of mental illness detection [6,8,10,12], the dataset is split into a ratio of 80:20 in this study. The class-wise distribution of documents related to the top and bottom three classes of Reuters and 20 Newsgroup is presented in

Table 6. Class-wise distribution of documents of benchmark datasets.

Reuters dataset
Top 3 class name with training and testing documents	Class name	Training: Testing	Bottom 3 class name with training and testing documents	Class name	Training: Testing
	Earn	2877:1087		Sun-meal	1:1
	Acq	1650:719		Lin-oil	1:1
	Money-fx	538:179		Coconut-oil	1:2
20 Newsgroups dataset
Top 3 class name with training and testing documents	Class name	Training: Testing	Bottom 3 class name with training and testing documents	Class name	Training: Testing
	rec.autos	485:109		soc.religion.christian	58:12
	rec.sport.hockey	476:124		alt.atheism	45:8
	rec.motorcycles	475:123		talk.politics.misc	43:9

. Preliminary analysis reveals that all the three datasets include uneven distribution of samples among different classes.

After applying the steps of preprocessing, the feature vectors for each dataset are created using the TF-IDF vectorizer. Then, the performance of the four RandNNs is observed with the ModCHI-based FS technique by systematically varying the value of selected features between 10 to 1500.

Table 7. Performance of RandNNs with ModCHI for the Reuters dataset.

Performance with ELM
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.9	0.91	0.92	0.93	0.94	0.94	0.95	0.94	0.94	0.94
Rmicro	0.33	0.43	0.56	0.68	0.69	0.68	0.7	0.71	0.72	0.73
Fmicro	0.48	0.58	0.70	0.79	0.80	0.79	0.81	0.81	0.82	0.82
HL	0.009	0.008	0.007	0.007	0.006	0.006	0.005	0.005	0.004	0.004
Performance with RVFLN
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.94	0.94	0.95	0.94	0.95	0.93	0.95	0.94	0.94	0.94
Rmicro	0.38	0.47	0.67	0.71	0.77	0.78	0.79	0.79	0.81	0.81
Fmicro	0.54	0.63	0.79	0.81	0.85	0.85	0.86	0.86	0.87	0.87
HL	0.008	0.007	0.006	0.004	0.004	0.004	0.003	0.003	0.003	0.003
Performance with KRVFLN
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.94	0.94	0.95	0.96	0.96	0.97	0.96	0.97	0.96	0.97
Rmicro	0.4	0.47	0.68	0.72	0.78	0.79	0.8	0.81	0.8	0.81
Fmicro	0.56	0.63	0.79	0.82	0.86	0.87	0.87	0.88	0.87	0.88
HL	0.008	0.007	0.006	0.006	0.004	0.004	0.004	0.003	0.003	0.003
Performance with BLS
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.94	0.94	0.95	0.95	0.97	0.97	0.98	0.98	0.97	0.98
Rmicro	0.38	0.50	0.67	0.72	0.78	0.79	0.83	0.83	0.83	0.83
Fmicro	0.54	0.65	0.79	0.82	0.86	0.87	0.90	0.90	0.89	0.90
HL	0.008	0.007	0.006	0.004	0.004	0.002	0.002	0.003	0.002	0.002

,

Table 8. Performance of RandNNs with ModCHI for the 20 Newsgroups dataset.

Performance with ELM
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.79	0.81	0.84	0.87	0.89	0.91	0.91	0.92	0.91	0.91
Rmicro	0.19	0.28	0.32	0.36	0.48	0.55	0.60	0.61	0.62	0.65
Fmicro	0.31	0.42	0.46	0.51	0.62	0.69	0.72	0.73	0.74	0.76
HL	0.05	0.04	0.04	0.04	0.04	0.04	0.04	0.03	0.03	0.03
Performance with RVFLN
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.83	0.85	0.89	0.92	0.92	0.92	0.91	0.90	0.90	0.89
Rmicro	0.23	0.33	0.34	0.39	0.54	0.59	0.62	0.62	0.62	0.68
Fmicro	0.36	0.48	0.49	0.55	0.68	0.72	0.74	0.73	0.73	0.77
HL	0.04	0.03	0.03	0.03	0.02	0.02	0.02	0.02	0.02	0.02
Performance with KRVFLN
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.82	0.86	0.89	0.92	0.92	0.93	0.93	0.93	0.94	0.94
Rmicro	0.25	0.37	0.38	0.40	0.55	0.60	0.64	0.67	0.67	0.67
Fmicro	0.38	0.52	0.53	0.56	0.69	0.73	0.76	0.78	0.78	0.78
HL	0.03	0.03	0.03	0.02	0.02	0.02	0.02	0.02	0.02	0.02
Performance with BLS
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.84	0.87	0.9	0.93	0.94	0.95	0.96	0.96	0.96	0.96
Rmicro	0.24	0.36	0.39	0.41	0.57	0.65	0.68	0.71	0.7	0.70
Fmicro	0.37	0.51	0.54	0.57	0.71	0.77	0.80	0.82	0.81	0.81
HL	0.04	0.04	0.03	0.03	0.02	0.02	0.02	0.02	0.02	0.02

and

Table 9. Performance of RandNNs with ModCHI for the Reddit Mental Health Dataset.

Performance with ELM
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.74	0.78	0.79	0.81	0.83	0.84	0.86	0.87	0.87	0.87
Rmicro	0.34	0.39	0.41	0.45	0.47	0.49	0.52	0.55	0.57	0.59
Fmicro	0.47	0.52	0.54	0.58	0.60	0.62	0.65	0.67	0.69	0.70
HL	0.08	0.08	0.08	0.07	0.07	0.07	0.06	0.06	0.06	0.06
Performance with RVFLN
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.77	0.79	0.79	0.81	0.83	0.85	0.86	0.88	0.89	0.89
Rmicro	0.35	0.39	0.44	0.48	0.53	0.56	0.60	0.61	0.63	0.63
Fmicro	0.48	0.52	0.56	0.60	0.63	0.65	0.68	0.69	0.74	0.74
HL	0.08	0.08	0.07	0.07	0.07	0.07	0.07	0.06	0.06	0.06
Performance with KRVFLN
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.78	0.79	0.81	0.83	0.84	0.85	0.87	0.89	0.92	0.92
Rmicro	0.35	0.38	0.45	0.49	0.54	0.57	0.61	0.63	0.65	0.63
Fmicro	0.48	0.51	0.58	0.62	0.66	0.68	0.72	0.74	0.76	0.75
HL	0.08	0.08	0.07	0.07	0.07	0.06	0.06	0.06	0.06	0.06
Performance with BLS
Evaluation Metrics	Feature Size (k)
	10	20	40	100	300	500	700	900	1000	1500
Pmicro	0.79	0.80	0.82	0.84	0.86	0.88	0.90	0.92	0.92	0.92
Rmicro	0.36	0.39	0.45	0.50	0.55	0.57	0.62	0.66	0.64	0.66
Fmicro	0.49	0.52	0.58	0.63	0.67	0.69	0.73	0.77	0.75	0.77
HL	0.08	0.08	0.08	0.07	0.07	0.07	0.06	0.06	0.06	0.06

present the results for P_micro, R_micro, F_micro, and HL values produced by four RandNNs utilizing the ModCHI with various feature sizes for the Reuters, 20 Newsgroups, and Reddit datasets, respectively. The best value observed related to each criterion for each RandNN is highlighted as bold in the tables.

By analyzing the results depicted in Table 7, it is observed that for ELM the highest P_micro value of 0.95 is achieved with a feature size of 700, the highest R_micro value of 0.73 is achieved with a feature size of 1500, and the highest F_micro value of 0.82 and the smallest HL value of 0.004 are achieved with feature sizes of 1000 and 1500 both. Similarly analyzing the outcomes of RVFLN with different k values of ModCHI, it is observed that the highest P_micro value of 0.95 is achieved with a feature size of 700, and the highest R_micro value of 0.81 and the highest F_micro value of 0.87 are achieved with feature sizes of 1000 and 1500 both. In contrast, the smallest HL value of 0.003 is achieved with feature sizes ranging from 700 to 1500. For KRVFLN, the highest P_micro value of 0.97 is achieved with feature sizes of 900 and 1500, and the highest R_micro value of 0.81 and the highest F_micro value of 0.88 are achieved with feature sizes of 900 and 1500 both. In contrast, the smallest HL value of 0.003 is achieved with feature sizes ranging from 900 to 1500. Similarly analyzing the outcomes of BLS with different feature sizes of ModCHI in the Reuters dataset, it is observed that the highest P_micro value of 0.98 and the highest F_micro value of 0.90 are achieved with feature sizes of 700, 900, and 1500, and the highest R_micro value of 0.83 is achieved with feature sizes ranging from 700 to 1500. In contrast, the smallest HL value of 0.002 is achieved with the same feature sizes of 500, 700, 1000, and 1500.

As depicted in Table 8, for the 20 Newsgroup dataset, the highest P_micro value of 0.92 is achieved with a feature size of 900, the highest R_micro value of 0.65 and the highest F_micro value of 0.76 are achieved with a feature size of 1500, and the smallest HL value of 0.03 is achieved with feature sizes of 900, 1000, and 1500 for the ELM network. Similarly analyzing the outcomes of RVFLN with different k values of ModCHI for the 20 Newsgroups dataset, it is observed that the highest P_micro value of 0.92 is achieved with the feature sizes of 100, 300, and 500, and the highest R_micro value of 0.68 and the highest F_micro value of 0.77 are achieved with a feature size of 1500. In contrast, the smallest HL value of 0.02 is achieved with a feature size ranging from 300 to 1500. For KRVFLN, the highest P_micro value of 0.94 is achieved with feature sizes of 1000 and 1500, and the highest R_micro value of 0.67 and the highest F_micro value of 0.78 are achieved with the feature sizes of 900, 1000 and 1500. In contrast, the smallest HL value of 0.02 is achieved with the feature size ranging from 100 to 1500. Similarly analyzing the outcomes of BLS over the 20 Newsgroups dataset, it is observed that the highest P_micro value of 0.96 is achieved with feature sizes ranging from 700 to 1500, the highest F_micro value of 0.70 is achieved with the feature size of 1500, the highest R_micro value of 0.82 is achieved with a feature size of 900, and the smallest HL value of 0.02 is achieved with the feature sizes ranging from 300 to 1500.

Analyzing the outcomes of four RandNNs for Reddit Mental Health Dataset, depicted in Table 9, it is observed that, the highest P_micro value of 0.87 is achieved with feature sizes of 900, 1000, and 1500. In addition, the highest R_micro value of 0.59 and the highest F_micro value of 0.70 are achieved when the feature size is 1500, and the smallest HL value of 0.06 is achieved with the feature sizes ranging from 700 to 1500 for the ELM network. Similarly analyzing the outcomes of RVFLN for Reddit dataset, it is observed that the highest P_micro value of 0.89, the highest R_micro value of 0.63, and the highest F_micro value of 0.74 are achieved with feature sizes of 1000 and 1500 both, whereas the smallest HL value of 0.06 is achieved with the feature sizes ranging from 900 to 1500. For KRVFLN, the highest P_micro value of 0.92 is achieved with the feature sizes of 1000 and 1500, and the highest R_micro value of 0.65 and the highest F_micro value of 0.76 are achieved when the feature size is 1000. In contrast, the smallest HL value of 0.06 is achieved with the feature size ranging from 500 to 1500. Similarly analyzing the outcomes of BLS for Reddit dataset, it is observed that the highest P_micro value of 0.92 is achieved with feature sizes ranging from 900 to 1500, and the highest F_micro value of 0.77 and the highest R_micro value of 0.66 are achieved with the feature sizes of 900 and 1500. In contrast, the smallest HL value of 0.06 is achieved with the feature sizes ranging from 700 to 1500.

After conducting a similar type of analysis on Table 7, Table 8 and Table 9 for each RandNN with 10 different k values of ModCHI across four evaluation criteria, it is found that a particular feature size does not give the best result for all performance measures. Thus, it points to a sensible situation of forming a more reliable decision considering multiple criteria. Further, selecting the optimal feature size for each RandNN is addressed as a Multi Criteria Decision Making problem, and TOPSIS is implemented to rank 10 alternates of each RandNN with respect to four criteria.

Table 10. TOPSIS ranking of different k values of ModCHI for each RandNN.

Dataset	Type of RandNN	Feature Size (k)
		10	20	40	100	300	500	700	900	1000	1500
Reuters	ELM	10	9	8	7	5	6	4	3	2	1
	RVFLN	10	9	8	7	5	6	3	4	1	2
	KRVFLN	10	9	8	7	6	4	5	1	3	2
	BLS	10	9	8	7	6	4	1	5	3	2
20 Newsgroup	ELM	10	9	8	7	6	5	4	3	2	1
	RVFLN	10	9	8	7	6	5	2	3	4	1
	KRVFLN	10	9	8	7	6	5	4	3	1	2
	BLS	10	9	8	7	6	5	4	1	2	3
Reddit	ELM	10	9	8	7	6	5	4	3	2	1
	RVFLN	10	9	8	7	6	5	4	3	1	2
	KRVFLN	10	9	8	7	6	5	4	3	1	2
	BLS	10	9	8	7	6	5	4	1	3	2

presents the TOPSIS ranking of all the feature sizes for each RandNN for all the three datasets. The optimal feature size selected by TOPSIS for each RandNN related to each dataset is further specified in

Table 11. Optimal feature size selected by TOPSIS for each RandNN.

Models	Reuters	20 Newsgroups	Reddit
	Feature Size	Feature Size	Feature Size
ELM	1500	1500	1500
RVFLN	1000	1500	1000
KRVFLN	900	1000	1000
BLS	700	900	900

. Analyzing the results of this table, it is clearly observed that the minimal feature size is selected for BLS with respect to each dataset.

To assess the impact of activation functions on the overall performance of the RandNNs, the performance of each NN is next evaluated with three activation functions, such as sigmoid, hyperbolic tangent (Tanh), and rectified linear unit (Relu), considering the selected features obtained by TOPSIS-ModCHI.

Table 12. Assessing the impact of activation functions for each RandNN with optimal feature set selected using the TOPSIS-ModCHI-based FS technique.

Dataset: Reuters
Models	Activation functions
	Sigmoid				Tanh				Relu
	Pmicro	Rmicro	Fmicro	HL	Pmicro	Rmicro	Fmicro	HL	Pmicro	Rmicro	Fmicro	HL
ELM	0.94	0.73	0.82	0.004	0.93	0.71	0.80	0.004	0.93	0.72	0.81	0.004
RVFLN	0.94	0.81	0.87	0.003	0.92	0.78	0.84	0.003	0.92	0.78	0.84	0.003
KRVFLN	0.97	0.81	0.88	0.003	0.95	0.77	0.85	0.004	0.94	0.67	0.78	0.005
BLS	0.98	0.83	0.90	0.002	0.97	0.78	0.86	0.003	0.97	0.77	0.85	0.003
Dataset: 20 Newsgroups
Models	Activation functions
	Sigmoid				Tanh				Relu
	Pmicro	Rmicro	Fmicro	HL	Pmicro	Rmicro	Fmicro	HL	Pmicro	Rmicro	Fmicro	HL
ELM	0.91	0.65	0.76	0.03	0.90	0.62	0.73	0.03	0.89	0.64	0.74	0.03
RVFLN	0.89	0.68	0.77	0.02	0.87	0.66	0.75	0.03	0.87	0.65	0.74	0.02
KRVFLN	0.94	0.67	0.78	0.02	0.92	0.65	0.76	0.02	0.91	0.63	0.74	0.03
BLS	0.96	0.71	0.82	0.02	0.95	0.67	0.78	0.03	0.92	0.65	0.76	0.03
Dataset: Reddit
Models	Activation functions
	Sigmoid				Tanh				Relu
	Pmicro	Rmicro	Fmicro	HL	Pmicro	Rmicro	Fmicro	HL	Pmicro	Rmicro	Fmicro	HL
ELM	0.87	0.59	0.70	0.06	0.86	0.57	0.68	0.06	0.84	0.55	0.66	0.07
RVFLN	0.89	0.63	0.74	0.06	0.87	0.61	0.71	0.06	0.87	0.60	0.71	0.06
KRVFLN	0.92	0.65	0.76	0.06	0.90	0.63	0.74	0.06	0.90	0.62	0.73	0.06
BLS	0.92	0.66	0.77	0.06	0.91	0.66	0.76	0.07	0.90	0.65	0.75	0.06

shows the outcome of each RandNN with the optimal feature set with respect to the three activation functions. Analyzing the results for all the three datasets, it is observed that in all the three cases BLS with sigmoid activation function is producing better values for each criterion.

Table 13. Assessing the impact of hidden layer size for BLS.

Dataset: Reuters						Dataset: 20 Newsgroups					Dataset: Reddit
Evaluation Metrics	Hidden Layer Size					Hidden Layer Size					Hidden Layer Size
	100	200	323	400	500	100	200	320	400	500	100	200	309	400	500
Pmicro	0.95	0.96	0.98	0.98	0.98	0.91	0.94	0.96	0.95	0.97	0.89	0.90	0.92	0.92	0.93
Rmicro	0.79	0.82	0.83	0.83	0.82	0.66	0.68	0.71	0.72	0.71	0.62	0.64	0.66	0.65	0.64
Fmicro	0.86	0.88	0.90	0.90	0.89	0.76	0.78	0.82	0.81	0.81	0.73	0.74	0.77	0.76	0.75
HL	0.003	0.003	0.003	0.003	0.003	0.03	0.02	0.02	0.02	0.02	0.05	0.05	0.06	0.06	0.06

shows the outcome of BLS utilizing the sigmoid activation function and the optimal feature set selected by the hybrid FS technique with five different hidden layer sizes to assess the impact of hidden layer size on its performance. It has been observed from Table 12 that, for most of the criteria, the better values are observed by considering the hidden layer size 323 for Reuters, 320 for 20 Newsgroups, and 309 for the Reddit dataset that was set using the approach specified in Equation (9).

Further analyzing the comparative performance of the four RandNNs with respect to the four criteria represented in Figure 6, Figure 7 and Figure 8, it is also clear that BLS with the least number of features selected by TOPSIS-ModCHI performs better as compared to the other three RandNNs for all three datasets. The true positive document selection of TOPSIS-ModCHI-based FS for all four RandNNs depicted in Figure 9, Figure 10 and Figure 11 for the three datasets also clearly shows that BLS is able to choose a high proportion of true positive documents with fewer attributes compared to the remaining three NNs.

Further, the statistical significance of the BLS model over KRVFLN, RVFLN, and ELM classifiers is evaluated using a resampled paired t-test [41]. In this technique, the splitting procedure used in the hold out method for the dataset is repeated for n times. In each iteration, the two models A and B used for comparison are trained on the training set and evaluated on the testing set. The performance difference of the two models is calculated for each iteration, resulting in n performance differences. Subsequently, the t statistic is computed using the following equation:

$$\mathrm{t}={\displaystyle \frac{\overline{\mathrm{p}}\surd \mathrm{n}}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}(}{\mathrm{P}}^{\left(\mathrm{i}\right)}-\overline{\mathrm{p}}{)}^2/(\mathrm{n}-1)}}$$

(14)

with n − 1 degrees of freedom, under the null hypothesis that models A and B perform equally. Here, P⁽ⁱ⁾ computes the difference between the model performance in the ith iteration, and ${\mathrm{P}}^{\left(\mathrm{i}\right)}={\mathrm{P}}_{\mathrm{A}}^{\left(\mathrm{i}\right)}-{\mathrm{P}}_{\mathrm{B}}^{\left(\mathrm{i}\right)}$ and $\overline{\mathrm{p}}$ represents the average difference between the classifier performance.

$$\overline{\mathrm{p}}={\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\begin{array}{c}{\mathrm{p}}^{\left(\mathrm{i}\right)}\\ \end{array}.$$

(15)

After calculating the t statistic, the next step is to compute the p value and compare it to the chosen significance level α. If the p value is less than α, we reject the null hypothesis and conclude that there is a significant difference between the two models. In this study, the paired t-test is conducted using Fmicro as the evaluation criterion with the significance level of 0.05.

Table 14. Paired t-test statistics.

Dataset	BLS/KRVFLN		BLS/RVFLN		BLS/ELM
	t-Statistics	p	t-Statistics	p	t-Statistics	p
Reuters	6.19	7.99 × 10−5	14.23	8.9 × 10−8	33.41	4.74 × 10−11
20 Newsgroup	20.12	4.29 × 10−9	14.45	7.77 × 10−8	28.94	1.70 × 10−10
Reddit	8.51	6.72 × 10−6	15.46	4.33 × 10−8	33.46	4.67 × 10−11

presents the t statistics and the corresponding p-value obtained from the resampled paired t-test with the n value set to 10. The results shown in Table 14 indicate that, the p-value is less than the predetermined significance level of α = 0.05 for all the three datasets. In each case, the p-value falls below the significance threshold, allowing us to reject the null hypothesis. This suggests a significant difference in the performance of the BLS model compared to the KRVFLN, RVFLN, and ELM models.

5. Discussion

In light of the fact that social media data might offer valuable insights into individual well-being, this study has investigated the use of four RandNNs to forecast mental illness. In order to enhance the performance of the RandNNs when processing text documents with uneven class distributions, a hybrid TOPSIS-ModCHI-based FS technique is recommended in the classification framework. As an improved version of chi-square, ModCHI emphasizes the selection of features with high class significance and low repetition. The ideal feature size k of ModCHI for each RandNN is not determined at random; instead, TOPSIS is used to figure out this value. Using two benchmark multiclass text datasets and the publicly available Reddit Mental Health Dataset, the efficacy of the proposed model is assessed through experimentation in various circumstances.

Analyzing the P_micro, R_micro, F_micro, and HL values depicted in Table 7, Table 8 and Table 9 related to the four RandNNs using ModCHI by systematically varying its k value between 10 and 1500, it is observed that a single k value of ModCHI does not yield the best result for all performance measures. Further approaching the selection of the optimal k value of ModCHI for each RandNN as a MCDM problem, TOPSIS is applied to rank the 11 feature sizes with respect to four criteria. As evident from the values depicted in Table 10, it is found that in the case of the Reuters dataset, the ideal feature sizes of ModCHI for ELM, RVFLN, KRVFLN, and BLS are 1500, 1000, 900, and 700, respectively. For the 20 Newsgroups dataset, the ideal feature sizes of ModCHI for ELM, RVFLN, KRVFLN, and BLS are 1500, 1500, 1000, and 900, respectively. In case of the Reddit dataset, the ideal feature sizes of ModCHI for ELM, RVFLN, KRVFLN, and BLS are 1500, 1000, 1000, and 900, respectively. Upon further examination of the ideal feature size of ModCHI for each RandNN, as shown in Table 11, it comes to light that the proposed hybrid FS approach selects the fewest features for BLS with respect to each dataset.

The comparative performance of the four RandNNs with least number of features selected by TOPSIS-ModCHI with three different activation functions, as shown in Table 12, reveals that in the Reuters dataset, the BLS model using sigmoid function produces an improvement of 4.25%, 4.25%, and 1.03% in P_micro value; 13.69%, 2.46%, and 2.46% in R_micro value; 14.6%, 3.44%, and 2.27% in F_micro value; and 50%, 33.33%, and 33.33% in HL value as compared to ELM, RVFLN, and KRVFLN, respectively. For the 20 Newsgroups dataset, the BLS model with sigmoid outperforms the ELM, RVFLN, and KRVFLN by 5.49%, 7.86%, and 2.12% in the P_micro value; 9.23%, 4.41%, and 5.97% in the R_micro value; 7.89%, 6.49%, and 5.12% in F_micro; and 33.33%, 0%, and 0% in the HL value. In the Reddit Mental Health Dataset, with a sigmoid function, the BLS produces same P_micro value as that of KRVFLN with less selected features and an improved P_micro value by 5.74% and 3.37% as compared to ELM and RVFLN, respectively. It improves the R_micro value by 11.86%, 4.76%, and 1.53% as well as the F_micro value by 10%, 4.05%, and 1.31% as compared to ELM, RVFLN, and KRVFLN, respectively. Moreover, it yields the same HL value as RVFLN and KRVLN with less selected features, and it produces an improvement of 14.28% in the HL value when compared to the ELM model. It is also evident from the overall comparison of the four RandNNs based on four criteria shown in Figure 6, Figure 7 and Figure 8 that the BLS with the fewest features chosen by TOPSIS-ModCHI outperforms the other three RandNNs for each of the three datasets.

From Table 13, it can be depicted that the BLS model gives the best performance with a feature size of 700 and a hidden layer node size of 323 in the Reuters dataset with a P_micro value of 0.98, R_micro value of 0.83, F_micro value of 0.90, and HL value of 0.003. For the 20 Newsgroups dataset, the BLS model gives the best performance with a feature size of 900 and hidden layer node size of 320 with a P_micro value of 0.96, R_micro value of 0.71, the F_micro value of 0.82, and HL value of 0.02. In case of Reddit Mental Health Dataset, the BLS model gives the best performance with a feature size of 900 and hidden layer node size of 309 with a P_micro value of 0.92, R_micro value of 0.66, F_micro value of 0.77, and HL value of 0.02.

As evidenced by the classification results in Table 12 and Table 13, the BLS consistently outperforms the other three RandNNs across multiple performance metrics, particularly in handling the complexities of multiclass and imbalanced mental health datasets. Further, the statistical analysis presented in Table 14 underscores BLS’s capability in handling multiclass unbalanced mental health datasets with minimal features, offering a promising approach for a scalable and efficient mental health analysis. Unlike ELM and RVFLN, which use a single-pass training process, the combination of broad feature mapping with enhancement nodes helps to improve the fitting ability of BLS compared to other RandNNs. The layered structure allows BLS to capture the nonlinear relationships in complex and sparse textual data in a more efficient way. In summary, the comparative analysis reinforces that BLS delivers a promising trade-off among performance, computational cost, and generalization, positioning it as a preferred choice for efficient mental health detection.

6. Conclusions

In conclusion, this study highlights the potential of leveraging ML in mental health analysis, particularly through the application of RandNNs. The exploration of the four RandNNs, such as BLS, RVFLN, KRVFLN, and ELM, coupled with the innovative TOPSIS-ModCHI-based FS technique, brings new insights in handling multiclass text data with unbalanced class distributions, which are common in mental health datasets. The integration of ModCHI for FS by focusing on the distribution of relevant documents across classes with the TOPSIS for optimal feature size selection showcases the potential of combining statistical and decision making methodologies in enhancing model robustness by ensuring better FS and improved classification accuracy. The experimental results and discussion presented related to two benchmark multiclass text datasets and the publicly available Reddit Mental Health Dataset clearly reveal that, among the four RandNNs, BLS emerged as the most effective model providing the highest P_micro, R_micro, and F_micro values and the lowest HL value surpassing the other models. This underscores BLS’s capability in handling multiclass unbalanced mental health datasets with minimal features, offering a promising approach for scalable and efficient mental health analysis. The insights gained from this research could help mental health professionals leverage social media data effectively to understand and predict mental health trends.

However, there are some additional aspects that remain unexplored in our research. First, we performed all the experiments with a single mental health dataset, considering eight categories of mental disorders. In future work, the model could be trained and tested using a broader range of social media datasets and additional mental health categories to improve its generalizability and robustness. Although we have tested our model with a multiclass unbalanced dataset, it can be extended by exploring the performance of RandNNs for multilabeled text classification. This would be valuable in scenarios where a single social media post may indicate multiple co-occurring mental disorders. The performance of BLS could also be enhanced by integrating suitable optimization techniques in its weight updating phase as well as the FS process. RandNN-based ensemble models could be developed and explored for mental health analysis. Again, the current work focuses on achieving high predictive performance; it does not yet address the interpretability of the models, which is critical for adoption in clinical settings. Future work will focus on developing a deployable solution in the form of a web-based dashboard or API for use by the clinicians. Additionally, involving domain experts in model evaluation and interpretation will be a key step toward ensuring that predictions align with clinical knowledge and practice.