Systematic sentiment analysis in mental health datasets is crucial for researchers investigating depression and anxiety, utilizing data sourced from both wearable devices and social media platforms. While some studies have employed transformer models for depression detection, their findings remain limited due to the lack of comprehensive evaluations across diverse datasets, which reduces the generalizability of the results. Furthermore, insufficient reporting on false positive and false negative detection methods weakens the robustness of their conclusions (Verma et al., 2023). Comparative analyses of depression and suicide detection using machine learning and transformer models have also overlooked the effects of dataset bias and class imbalance on model performance. The absence of extensive testing across multiple social media platforms restricts the practical applicability of these approaches, highlighting a significant gap in current research (Bokolo and Liu, 2024).

The research evaluation was limited by a small data sample and unreliable data quality, which decreased the universal applicability of the results regarding wearable technology-based HRV prediction of mental health and overall wellness. Such omissions regarding participant characteristics, including BMI, result in weak generalizations of the study outcomes (Coutts et al., 2020). The proposed deep learning method for depression intensity measurement on social media did not address the negative effects of noisy data on model precision. The model displayed limited ability to recognize diverse populations because the dataset was not diverse (Ghosh and Anwar, 2021). Another study suggested using a hybrid deep learning system to detect depression but failed to examine the security issues related to the vulnerability of social media datasets. The methodology did not account for the fact that language usage varies across social media sites, which created limitations for the model's practical application (Zogan et al., 2022). A deep learning system to identify mental illness from social media did not account for the broad variability of mental symptoms that influence its precision level. The model considered only text information and did not extract corresponding context information regarding user interaction or media post content (Kim et al., 2020). For predicting the risk of depression, deep models were used; however, they failed to consider that multivariable regression could miss identifying non-linear associations between variables in the model. The performance of this model could be degraded by the lack of diversity in real-world data (Baek and Chung, 2020). The study on machine learning and deep learning diagnosis techniques did not include data bias analysis, which could affect the fairness of the model. The lack of clarity in the model process negatively impacted transparency, a crucial aspect of designing mental health applications (Kasanneni et al., 2025).

Important ethical considerations arise when using personal data in a study on predicting mental health consultations from social media posts. The linguistics-based model may have limitations in precision when handling various populations digitally across different platforms (Saha et al., 2022). The final part of the work involved sentiment analysis through the integration of the Bi-LSTM and BERT models in depression prediction. However, it did not consider textual elements such as sarcasm and irony. Furthermore, the model lacks the capacity to perceive non-verbal cues and multimedia elements, as it relies solely on textual information (Boer et al., 2021).

Transformers possess exceptional power and capability in capturing long-range dependencies and contextual relationships within sequential data through their self-attention mechanisms. This makes them particularly effective for sentiment analysis tasks, where understanding complex language patterns and context is crucial for accurately detecting sentiment polarity and intensity. In literature, the research model faced trust issues because it solely used texts without integrating multiple types of evidence, and the authors withdrew their work (Zeberga et al., 2022). Another study employed transformer-based machine learning for counseling conversation analysis, although it failed to address ethical issues related to the use of therapy data. Prediction accuracy could be improved by incorporating non-verbal cues, as text-based analysis often provides insufficient information (Imel et al., 2024). The assessment of eating disorder-related tweets using machine learning methods alongside BERT models yielded inadequate results because the system design overlooked the diverse expression methods within the text data. Text-only data lacked contextual information about user engagement and multimedia content, which could be vital for understanding the problem (Benítez-Andrades et al., 2022). A transformer-CNN hybrid model for cognitive behavioral therapy assessment in psychological testing did not address the performance inefficiency and resource requirements associated with combining multiple models. The evaluation failed to explore unstructured behavioral and contextual cues, which could enhance both diagnosis accuracy and treatment efficiency (Ankalu Vuyyuru et al., 2023). The analysis of machine learning algorithms alongside deep learning for mental health diagnosis from social media platforms failed to resolve data quality and imbalance issues. Using textual data alone prevented healthcare professionals from accessing multimodal features, which could enhance diagnostic efficiency (Chiong et al., 2021). The social media behavioral analysis system, PsychBERT, did not address potential ethical problems that could arise from using individual data. The text-based analysis approach had a potential drawback because it could not detect important contextual signals that multimedia elements and user activity would provide (Vajre et al., 2021).

A transformer-based deep learning model examined suicidal emotions on social media yet failed to address subtle or hidden expressions because it reduced the system's accuracy level. The approach used only text as its basis while ignoring essential visual and user-related information cues (Kodati and Tene, 2023). The combination of explainable AI with machine learning analyzed Reddit wellbeing but failed to consider linguistic differences and posting contexts, making it difficult to achieve accurate results. The sole reliance on text data prevented the system from discovering vital behavioral and multimodal information that could strengthen prediction accuracy (Thushari et al., 2023). The implementation of RoBERTa-Large and BERT in mental healthcare applications restricted their capacity to handle specialized vocabulary within this domain, thus affecting the system's accuracy. Applications could achieve better performance by incorporating non-verbal cues through text analysis alone (Wu et al., 2024). XAI transformer-based interpretation methods were developed to understand depressed and suicidal user tendencies, although researchers failed to address the difficult nature of detecting subtle expressions. The analysis of text alongside structured data failed to capture non-verbal signals together with situational context, which reduced the accuracy of the interpretation (Malhotra and Jindal, 2024).

In this study, data preprocessing is a crucial step to ensure the accuracy and reliability of sentiment analysis on mental health-related content. The raw text data collected from social media platforms contains a lot of noise, such as irrelevant words, special characters, numbers, and stop words, which can negatively impact the performance of Natural Language Processing (NLP) models (Wu et al., 2023).

Several preprocessing techniques are applied to the textual data to improve its quality (Ahmed et al., 2023). Initially, stop words S ϵ {s₁, s₂, ...., s₃, words w_j such as “the” and “is,” which do not carry any useful information in sentiment classification, are removed using function f_stop from dataset D, as defined in Equation 1. Table 2 provides a description of the symbols used in the equations for a deeper understanding.

Moreover, set C containing digits and characters c_k ϵ C, such as punctuation marks or emojis, are also eliminated using function f_char using Equation 2, for textual uniformity and minimum variations in the dataset.

After removing the components that might not be relevant, normalized texts are generated to make the content consistent and minimize the differences among word forms, using function f_norm, using Equation 3. This happens by converting all text to lowercase, so that when singling out noise words, they are standardized and independent of case sensitivity.

After that, we apply lemmatization using f_lemma to lower all words to their base forms to thereby increasing the model's efficiency by limiting redundant word variations. For example, words, such as “running,” “ran” are changed to “run,” where similar semantically words are treated equally, as in Equation 4. Lemmatization preserves the contextual meaning of words while reducing data dimensionality, which aids in better model generalization.

Another important step is that the NLP model must initially process the text coverage, which involves tokenization f_token, in which the text is previously divided into several words or subwords, allowing the NLP model to understand linguistic patterns, as defined in Equation 5. The model can capture syntactic and semantic relationships by breaking down sentences into meaningful terms.

Deep learning architectures, such as transformer-based models, are then used to process the tokens generated after preprocessing. Sentiment analysis becomes more efficient when the data are preprocessed, with the noise first removed, text structures standardized to facilitate further data processing, and only the most valuable features selected for classification. This structured approach enhances model performance in detecting sentiment patterns related to mental health concerns such as anxiety, depression, or emotional distress.

RoBERTa-Large (Robustly Optimized BERT Pretraining Approach) is an improved transformer-based model that enhances the performance of BERT by refining its pretraining strategy. RoBERTa-Large is based on the BERT architecture but does not include the Next Sentence Prediction (NSP) objective. It utilizes larger batch sizes and is trained on more data with dynamic masking, developed by Facebook AI. RoBERTa-Large's performance is rich in a variety of tasks centered on NLP, ranging from sentiment analysis to classification to mental health detection, as it has 24 transformer layers, 16 attention heads, and 355 million parameters (Youngmin et al., 2024). Due to its ability to leverage fine-grained language representations, it is ideal for analyzing complex textual data (e.g., user-generated content about mental health discussions), as proposed by the model architecture defined in Figure 4.

Efficiently learning an Encoder that Classifies Token Replacements Accurately (ELECTRA) uses a pre-training approach called Replaced Token Detection (RTD). ELECTRA does not predict masked tokens; instead, it learns to distinguish masked tokens from plausible replacements that it generates using a small generator network, as shown in the working architecture defined in Figure 5. The ELECTRA model consists of two components: Generator G and Discriminator D. The generator is a small, masked language model (MLM) model that generates the masked tokens of a given input sequence. It produces mismatched versions of input but generates plausible replacements for masked tokens. The generator is trained to replace the tokens with replacements that would be generated by the discriminator, a larger transformer. ELECTRA is more efficient and effective than BERT, as it can evaluate each token in the sequence, not just the masked ones (Ahmed et al., 2024). Both the generator and the discriminator are trained jointly, but only the discriminator is utilized in downstream tasks, such as sentiment analysis.

For the experiments, the dataset is sourced from the Kaggle website, which contains a collection of mental health-related statements from various datasets, including the 3k Conversations Dataset for Chatbot, Depression Reddit Cleaned, Human Stress Prediction, and others. The dataset content is based on reviews generated by users of online platforms such as Reddit and Twitter but is annotated with one of the seven mental health statuses, including normal, depression, suicidal ideation, anxiety, stress, bipolar, and personality disorder. Each entry consists of a unique identifier, a text statement, and a corresponding mental health label. This dataset provides a substantial amount of data for machine learning models to utilize for sentiment analysis and chatbot development, which can aid in early detection and support for mental health conditions.

For the comparative analysis, several state-of-the-art deep learning models are considered for detecting complex patterns.

The proposed model was developed and fine-tuned using widely adopted deep learning and NLP libraries, including PyTorch and the Hugging Face Transformers framework. Data handling and preprocessing were performed using Pandas and NumPy, while scikit-learn was utilized for evaluation metrics. Visualization and interpretability were supported through Matplotlib and Seaborn, as summarized in Table 3. The training and evaluation were conducted on a GPU-enabled environment using Google Colab to ensure efficient computation.

The dataset is compiled from comments shared by students, focusing on their mental health experiences and the impact of education and physical activities on their wellbeing. Students were encouraged to express their thoughts on how various activities impacted their mental state in terms of stress, anxiety, depression, and overall psychological resilience. The responses reflect a wide range of emotions and sentiments related to student activities, academic pressures, and personal challenges. These textual responses are structured and labeled in the dataset, serving as input for sentiment analysis and mental health assessment. Such data can be used by AI models to identify patterns and trends related to the mental wellbeing of students.

The pie chart in Figure 6 provides crucial insights into the dataset's composition regarding various mental health statuses. The dataset comprised seven classes with the following distribution: normal (16,351 samples), depression (15,404 samples), suicidal ideation (10,653 samples), anxiety (3,888 samples), bipolar disorder (2,877 samples), stress (2,669 samples), and personality disorder (1,201 samples). The most frequently occurring labels are “depression and suicidal tendencies,” suggesting that the collected data are more likely to involve these labels.

Significantly also represented is anxiety about its widespread presence in mental health discussions. On the other hand, possible class imbalance is higher in categories like bipolar disorder and personality disorder, and such classes are rare, whereas the described category is common. The dataset distribution also matches with that of real-life trends, where depression and anxiety are more prevalent on social media, and the prevalence of disorders such as bipolar and personality disorders might be less represented, potentially because of stigma and less self-reporting on public forums. For robust NLP models for sentiment analysis, the distribution of labels must be known. It aids in determining the need for techniques to balance feature engineering methods and evaluation metrics to achieve favorable model performance across all categories of mental health. The skewed distribution also suggests that the model should be evaluated carefully, considering a precision–recall tradeoff to ensure that predictions for the minority class are reliable.

The correlation heatmap in Figure 7 provides insights into the relationships between the various textual features in the dataset. Statement length (r = 0.479) and number of words (r = 0.466) exhibit a moderate positive correlation with the status variable, indicating that different mental health conditions may require longer statements. Moreover, we observe that the feature statement length and the number of words clearly correlate (r = 0.995), as longer statements obviously contain more words. Average word length, however, has a very weak correlation with all other features, suggesting that there is little difference in word complexity within mental health conditions. This proves that textual attributes can be utilized for sentiment analysis in mental health classification.

The word clouds in Figure 8, from the provided visualizations, indicate the words most frequently used in relation to mental health statuses. The word clouds provide an overall linguistic representation of the user-generated content's depiction of mental health status.

Overall, in Figure (a), they feel, want, know, think, live, and other words that dominate this word cloud represent the entire dataset. These words fit within the sphere of introspective and emotional discussions about the disease. “Help” and “work” are also common words that indicate how external support and occupational stress may play a role in shaping the language of mental health.

Personality Disorder in Figure (b): Words such as “feel,” “life,” “help,” and “think” are scattered within the personality disorder category. This means that people in this category tend to consider their emotions, relationships, and the self. Help implies that many call for external aid or help.

Anxiety in Figure (c): The word cloud related to anxiety includes “feel,” “think,” “know,” “life,” and “anxiety.” The overthinking that comes through thinking and knowing suggests that anxiety is prominent here. Through the repetition of feeling, the pronounced repetition of feeling highlights how anxious thinking is associated with emotional distress, implying the need for reassurance or professional assistance.

In this normal status category, as shown in Figure (d), the word cloud displays words such as “feel,” “life,” “think,” and “know,” and does not carry the heavy emotional weight of the other categories. This category comprises statements that, in most cases, are not particularly emotional, yet they still involve a neutral or positive discussion. This implies that their terms are balanced and do not have any significant mental health concerns or conversational patterns.

Depression in Figure (e): Words such as “feel,” “life,” “want,” “help,” and “think” occur most in the word cloud of the depression-related words. Life and want are two things that indicate existential feelings, such as longing or hopelessness. The prominence placed on help highlights the importance of support systems for individuals who are depressed.

Suicidal ideation in Figure (f): Words in the suicidal ideation category are prominent, including “want” and “life,” and “know” and “feel,” indicative of distressing thoughts and existential, and feeling words. “End” is also used in these statements, suggesting that there is emotional chaos of the worst kind imaginable. This is a pattern of behavior that requires mental health interventions urgently, by people saying such things. Included in the stress-related word cloud in Figure (g) are the words “stress,” “work,” “time,” and “life,” which all relate to it. This implies that work and time, if they are dominant, are primary factors leading to stress. Therefore, in addition to this, thinking and knowing indicate cognitive strain, where people think about and understand their stressful situations.

Bipolar in Figure (h): Among the most prominent words in the bipolar disorder category are “go,” “time,” “work,” “feel,” and “help.” Bipolar mood swings are often signaled by fluctuating energy levels and erratic thought processes, both of which point to “go” and “time.”

This analysis, which utilizes word clouds to explore the frequency of the most common words in a corpus, facilitates an understanding of the linguistic markers associated with different mental health conditions.

The proposed model was applied to the student mental health dataset, yielding reliable detection of key psychological patterns and risk factors. The findings highlight the model's potential in supporting early identification and intervention for students' wellbeing.

By examining confusion matrices that predict the mental health of students using ELECTRA, GRU, LSTM, and Bi-LSTM models, this study provides a clear understanding of the strengths and limitations of these approaches. These models were subsequently evaluated under various conditions, annotated as anxiety, bipolar, depression, normal, personality disorder, stress, and suicidal ideation, which show the predictive capabilities and inadequacies by means of rates of correct and incorrect classification for the model.

Table 6 presents various feature-based models applied in mental health cases to yield results. A comparative analysis of the proposed RoBERTa-Large model with existing models shows that our model achieves a 97% result in sentiment analysis of mental health, relying on word embedding's. Indeed, this model surpasses other approaches, such as the traditional LSTM, which utilizes HRV from wearable devices, as well as textual features from Twitter, to achieve results of 83% and 74% in earlier applications. Among more advanced models, such as BiLSTM+BERT and other models like MentalBERT and RoBERTa+BERT, which perform well on complex text data and social media interactions, the highest performance achieved is only up to 89% and 79%, respectively. For example, a more recent development, XLNet, which utilizes word embedding's for social text-based communication, was more effective than RoBERTa-Large, achieving an outcome of 92%. Based on the above, the RoBERTa-Large model is selected for its superior performance, as well as its effective handling of word embedding's and robust training, which likely corresponds to a deeper, more nuanced understanding of language context for mental health. This makes it a brilliant tool in the domain of mental health diagnostics and sentiment analysis.