An explainable AI framework integrating deep learning and large language model for student’s mental health

1 Introduction

Mental health disorders like anxiety, depression, and MCI are common among young adults, particularly within the age range of 18–30 years, especially among students. These conditions extensively affect cognitive performance, emotional stability, and overall mental wellbeing (Sawicki et al., 2026; Ben Simon et al., 2026; Aljebreen et al., 2026; Alharithi, 2026). Early detection of mental health disorders is critical. However, traditional mental health diagnosis is based on self-reported questionnaires and clinical interviews, which are particularly inefficient and often insufficient for detecting mental health disorders. In recent years, artificial intelligence and deep learning approaches have been explored for mental health prediction. The existing approaches apply Machine Learning (ML) models, Convolutional Neural Networks (CNN), and transformer models for mental health prediction. Few Techniques also try to use multimodal features such as text responses, physiological responses, and survey responses for better prediction accuracy (Nogay and Adeli, 2026; Nandan et al., 2026; Tutun et al., 2023). Even though these approaches have shown promising results, there are many challenges in effectively modelling mental health indicators.

Despite these advances, certain limitations still exist; most of the existing models are based on single-modal or limited multimodal data, which does not provide thorough understanding of the mental state of a person. Moreover, the absence of interpretability in Deep Learning (DL) models makes it difficult for these models to be used in clinical practice, as healthcare providers require transparent decision-making mechanisms (Sharma et al., 2025; Choudhury et al., 2026; Hochstenbach et al., 2026). Additionally existing models have poor generalization across different age groups. A major research gap exists in the limited focus on high-risk and vulnerable populations, who are often underrepresented despite having a higher susceptibility to mental health disorders. Furthermore, most existing systems focus only on prediction and fail to provide actionable recommendations. These issues highlight the need for a unified, interpretable, and robust framework that supports both prediction and decision-making in mental health examination.

To overcome these challenges, this research proposes an interpretable DL method for multimodal prediction of anxiety, depression, and MCI. The proposed framework utilizes demographic, cognitive, behavioral, psychological, and physiological features using a cross-domain feature learning method. The framework employs an MPDSA-CD for accurate classification and SHAP for interpretability. The proposed framework enhances accuracy, reliability, and interpretability for early detection and effective management of mental health disorders among young adults. The major contribution of the research includes:

• Proposes an MPDSA-CD method for robust multimodal and cross-domain representation learning, enabling accurate classification of mental health conditions and early cognitive decline detection in young adults aged 18–30, integrated with Kruskal-Wallis-based statistical feature selection to identify discriminative features and reduce redundancy.
• Introduce a CDFLN to effectively integrate heterogeneous data (demographic, behavioural, cognitive, and physiological), facilitating enhanced cross-domain feature interactions and representation of complex mental health patterns.
• Developed dense feature propagation with a self-attention mechanism to capture both intra- and inter-modal relationships, improving feature representation and classification capability across anxiety, depression, and MCI.
• Incorporated SHAP-based explainability with a cognitive risk assessment module to provide interpretable predictions and continuous severity scores, supporting clinically meaningful decision-making.

The rest of this research is organized as follows: Section 2 reviews the related work. Section 3 describes the proposed framework and its key components. Section 4 presents the experimental setup, evaluation metrics, and results. Section 5 concludes the study and outlines future research directions.

2 Literature survey

Atlam et al. (2025) suggested an Explainable Mental Health Disorders (EMHD) model for predicting autism-related mental disorders using ensemble learning and SHAP. Although the model offers high accuracy, it can only handle structured data and lacks multimodal deep feature learning. Hill et al. (2025) suggested a neural network model utilizing psychosocial and neuroimaging information and obtained promising performance in predicting mental health risks among adolescents. However, low interpretability and little advantage of multi-modal fusion hindered the models’ clinical benefits. Tutun et al. (2023) developed an AI-based Decision Support System (DSS) for diagnosing mental disorders with an accuracy of 89% by using fewer questions with the help of the Network Pattern Recognition (NEPAR) algorithm and machine learning techniques. However, there is a high dependency on questionnaire-based inputs and a lack of time analysis of mental health progression. Vandana et al. (2023) suggested a hybrid DL model that incorporates a textual CNN, audio CNN, and Bi-LSTM for depression detection. The model has high accuracy, but it lacks explainability, real-time, and integration of multimodal features. Khan and Alqahtani (2024) suggested a hybrid approach using machine learning models with BERT, TF-IDF, ANN, and SVM for depression detection using Twitter data. Though the approach has high accuracy, it is restricted to using textual data alone and does not consider multimodal data. Also, social media data may have noise in it. Zhai et al. (2025) suggested an ML model including XGBoost, Random Forest, and Logistic Regression to predict anxiety and depression among college students. Although it is effective, this method lacks interpretability, real-time prediction, and deep feature integration. Bendebane et al. (2023) developed a multi-class model based on DL for the detection of depression and anxiety through Twitter data. However, the model lacks interpretability, relies on the quality of labelled data, and is not effective in handling long-term temporal patterns of mental health. Mumenin et al. (2025) introduced a DDNet that is a stacked ensemble hybrid model consisting of Multilayer Perceptron (MLP), Stochastic Gradient Descent (SGD), CatBoost (CB) and Least Absolute Shrinkage and Selection Operator (LASSO) for depression detection with high accuracy. Though the model has high performance, it also has high complexity, poor scalability, and lacks data integration. Balraj and Nagaraj (2026) introduced Hy-OCNN-BiLSTM for depression prediction using social media usage. Although it is highly accurate, it is based on social media, is a privacy risk, and is not clinically validated. Almadhor et al. (2024) developed a Multi-class Adaptive Active Learning (MAAL) method for student anxiety prediction. However, the model involves a labeling process, which is costly and may lead to instability. The model relies on the choice of the initial sample. Recent deep learning approaches based on techniques as Convolution Neural Network (CNN), Recurrent Neural Networks, Transformer based architecture and multimodal frameworks have contributed to improve the prediction accuracy.

2.1 Problem statement

Mental health disorders such as anxiety, depression, and MCI are increasingly prevalent among young adults, yet accurate and early detection remains challenging. Existing methods are often limited by reliance on single-modal or weak multimodal data, which fail to capture the complex and heterogeneous nature of human mental states. Moreover, there is a lack of integrated systems that provide both accurate prediction and meaningful risk assessment and clinical support.

3 Proposed methodology

The proposed CDFLN is a network architecture designed to capture and enhance critical patterns within complex data through multiscale and cross directional feature extraction. This cross directional feature captures inter-modal dependencies through cross-directional interactions between different modalities. This leads to improved prediction accuracy.

The Multimodal Progressive dense self-attention for cross domain. MPDSA-CD is an advanced deep learning architecture incorporated to capture complex cross-domain interactions and long-range dependencies through progressive dense self-attention mechanism. Self-attention mechanism analyses complex features, Attention enables interpretability and contextual sensitivity by giving higher importance to features strongly correlated with mental health outcomes. To handle data across different sources and formats the model applies domain adaptation technique, aligning features learned from one domain with those from another. This helps in transferring knowledge across domain and maintaining high prediction accuracy despite varying data distributions. Furthermore, the SFOA optimization strategy was employed to enhance parameter optimization to reduce convergence instability, and improve model generalization. Model employs dense connections across progressive layers each layer receives input from all preceding layers, ensuring feature reuse. This integrated components collectively overcome limitations associated with conventional multimodal approaches, especially in feature representation, interpretability, and predictive robustness.

The proposed framework includes various forms of data, such as demographic information, cognitive information, behavioral information, and physiological information. It is then utilized for the prediction of anxiety, depression, and mild cognitive impairment in young adults. It is processed using statistical selection and cross-directional feature learning. A progressive dense self-attention network is then utilized for classification using metaheuristic optimization. SHAP is utilized for the interpretation of the information, allowing for accurate prediction and early support. Figure 1 explains the overall architecture of the proposed method.

3.1 Data collection

Datasets were collected from the Mindcare Hospital for Mental and Sexual Health, Navi Mumbai and Morris Lifesciences and Technologies, USA involving individuals aged 18–30 from both India and the USA. Two Dataset records each was counted for evaluation (Google Drive link, 2026). The study employed two independent datasets: one from India with 27,901 participant records and the other from the USA with 27,901 participant records. The total number of samples was 55,802, and the total number of multimodal features was 53, across different domains of demographics, cognition, behavior, physiology, and psychology. The dataset includes comprehensive multimodal parameters such as demographic and academic information, cognitive functioning metrics, neurocognitive decline indicators, mental health and emotional states, physiological and health markers, and behavioral and lifestyle patterns. Total 53 parameters from six domains are counted for the evaluation. The study captures aspects like memory, stress, emotional stability, sleep, and daily behavior. Data collection source is standardized clinical assessments, structured questionnaires, and physiological measurements to ensure reliability. Classification of data was done at three levels as normal, anxiety or depressed. Participants covering predefine criteria such as age group, diagnosis confirmation are included whereas participant with missing responses, unrelated medical values are excluded to maintain consistency of data.

3.2 Data preprocessing

Data preprocessing is a significant step to guarantee the quality and usability of the collected data. The raw data collected from various sources, including clinical, behavioral, and psychological data, have missing values, noise, and varying formats. Therefore, a systematically designed data preprocessing pipeline is used to transform raw data into a clean, usable format. The missing values are handled by using efficient imputation strategies (Peng et al., 2024). For numerical features, the mean or median value is used based on the distribution of the data. Similarly, for categorical features, mode values are used. In some cases, where complex dependencies are involved, sophisticated imputation strategies like K-Nearest Neighbor (KNN) are used.

The data normalization is performed to bring all numerical features into a common range, which helps in efficient and stable learning. Min-Max normalization is used, which is defined as Equation 1.where x is the original value, and x ′ is the normalized value within the range [0,1]. This ensures that features with large values do not control the learning process. For handling the categorical data, encoding is used. For the ordinal data, label encoding is used, whereas for nominal data, one-hot encoding is used to avoid the concept of order. For the textual data and survey data, preprocessing is used. The preprocessing involves tokenization, lowercase, and removal of unwanted characters. This converts the data into numerical data, which is used for deep learning. In addition, data harmonization is also used. It ensures the data is consistent across the multimodal data. Furthermore, noise removal and outlier handling are also used.

x ′ = x − x min x max − x min (1)

3.3 Kruskal-Wallis test for statistical feature selection

After preprocessing, to determine the most discriminative features, the Kruskal-Wallis (K-W) test is performed for the groups, i.e., anxiety, depression, and MCI (Johnson, 2022). The K-W test is used to check if the distributions of the features, for each feature X, are significantly different across the groups. The test statistic is given by Equation 2.where g is the number of groups, n i is the number of samples in the group i, N = ∑ i = 1 g n i is the total number of samples, and R ¯ is the average rank of the feature X in group i, Alternatively, K expressed in terms of pairwise group differences is given in Equation 3.

H = 12 N N + 1 ∑ i = 1 g n i R i ¯ − N + 1 2 2 (2)

H = 12 N N + 1 ∑ i < j n i n j R i ¯ − R j ¯ 2 (3)

The feature is preserved if the associated p-value for the given value of H is less than the predetermined level of significance, denoted by α (for example, 0.05). This process of selection helps in reducing dimensionality, increasing the efficiency of the model, and ensuring that the relevant behavioral and cognitive factors are used in the subsequent predictive models.

3.4 Cross-directional feature learning network for feature extraction

The pre-processed selected multimodal features are further transformed into discriminative representations by using the CDFLN, which combines feature learning and cross-modal interaction. Assume that each sample is represented as S i = x i m m = 1 M , where x i m represents the input feature vector of the m t h modality (Zheng et al., 2023). Feature learning for individual modality-specific branches is done by using non-linear transformations to learn high-level representations, as given in Equation 4.where W m and b m are learnable parameters and σ i is an activation function. To ensure stable distribution of features and reduce modality-specific bias, an Adaptive Layer Normalization Unit (ALNU) is applied, as given in Equation 5.

f i m = ϕ m x i m = σ W m x i m + b m (4)

f ∼ m = A L N U f i m (5)

Subsequently, cross-directional feature learning is performed to capture inter-modal dependencies through pairwise interactions between modalities is given in Equation 6.where W m n represents cross-modal interaction weights and ⨂ represents element-wise multiplication. The final feature representation obtained by aggregating modality-specific and cross-modal features is given in Equation 7.

Z i m , n = f ∼ i m ⨂ W m n f ∼ i n , m ≠ n (6)

F i = c o n c a t f ∼ i 1 , f ∼ i 2 , . . . , f ∼ i M + ∑ m ≠ n Z i m , n (7)

This process allows the network to learn complex and non-linear features consistently by jointly learning intra-modal characteristics and inter-modal relationships, which improves the discriminative ability of features for accurate prediction of anxiety, depression, and mild cognitive impairment.

3.5 Multi-model progressive dense self-attention with cross-domain mental health classification

The improved multimodal feature representation F i, which is extracted from the CDFLN, is used to train the proposed network for the classification of mental health conditions. Unlike existing approaches that process features independently or lack interpretability, the proposed MPDSA-CD model effectively integrates dense feature reuse with attention-based contextual modelling, which is essential for handling heterogeneous multimodal and cross-domain data (Li et al., 2024). The multimodal feature representation provides heterogeneous information from various dimensions such as demographics, cognition, behavior, and physiology, facilitating comprehensive analysis and understanding of complex mental health conditions. In addition to multimodal fusion, the proposed framework explicitly incorporates cross-domain learning, where relationships between different domains (e.g., behavioral-cognitive, physiological-demographic) are jointly modelled. This allows the network to capture inter-domain dependencies, which are crucial for identifying subtle mental health patterns. Figure 2 explains the architecture of MPDSA-CD.

Traditional multimodal mental health prediction models use early fusion, late fusion or simple feature concatenation approaches. These methods incorporate information from several modalities, but do not necessarily account for complex cross-modality interactions among behavioral, cognitive, physiological, and demographic factors. Often, mental health disorders are multi-faceted and are a result of multiple interconnected factors, not a few individual ones. Thus, the use of learning modality-specific information alone may not be sufficient to adequately reflect psychological conditions. To overcome this limitation, the proposed Cross-Directional Feature Learning Network (CDFLN) has been designed to explicitly model interactions between heterogeneous modalities by cross-directional feature propagation. Compared with the conventional multimodal fusion techniques, which are mainly based on aggregation of features, CDFLN learns intra-modal representations and inter-modal dependencies. This allows the network to discover latent relationships, such as the combination of cognitive decline, emotional instability and behavioral changes, leading to more discriminative feature representations. Moreover, traditional deep learning architectures tend to lose information in each layer of the network when passing them to the next layer. To address this challenge, the proposed Multi-Model Progressive Dense Self-Attention Cross-Domain (MPDSA-CD) architecture incorporates the dense connectivity and self-attention mechanisms. Dense feature propagation encourages feature reuse and better gradient flow, minimizing potential loss of information. At the same time, the self-attention mechanism can learn from long-range contextual dependencies, and dynamically focus on clinically relevant indicators. This is to enable the model to concentrate on the most informative multimodal features and to suppress redundant information. Furthermore, mental health datasets are often encountered with multimodal feature distributions, and high-dimensional search spaces, which are difficult to optimize a model. Starfish Optimization Algorithm (SFOA) was incorporated to enhance the hyperparameter tuning and convergence behavior. It has been shown that SFOA can help improve exploration and exploitation in the parameter search process, thereby reducing the probability of local optimum and improving the robustness of the model, when compared with the traditional gradient based optimization method. This optimization approach is especially advantageous when the learning environment is multimodal, and there can be complicated interactions between features that yield very non-convex objective surfaces. In combination, the combined framework, CDFLN, MPDSA-CD and SFOA can learn modality specific features, cross-domain interactions, contextual dependencies and optimized decision boundaries. These features establish a solid foundation for better interpretability, robustness, and predictive performance than traditional multimodal deep learning methods.

Dense feature propagation is first performed to enable maximum information flow and feature reuse for effectively capturing intra- and inter-modal feature dependencies, as given in Equation 8.where H l denotes the output of the l − t h layer, W l and b l are learnable parameters, · represents the concatenation of features from all preceding layers. This dense connectivity enables cross-domain feature reuse, where information in one domain, such as cognition, can be used to support learning in another domain, such as behavior, thus avoiding loss of subtle information such as stress indicators and cognitive decline. The dense aggregation can be defined as Equation 9.where H d e n s e represents all modality-specific and cross-modality interactions and thus forms a unified feature space. To capture global contextual relationships across modalities, self-attention is used, which projects features to query, key, and value spaces as given in Equation 10.where Q, K, V are query, key, and value matrices, respectively. The attention weights are computed to learn the dependencies between multimodal features as expressed in Equation 11.where A is the normalized attention matrix and d is the scaling factor. The attention-enhanced feature representation can be obtained as H a t t = A V. Where H a t t highlights the most relevant multimodal features while suppressing redundant and less informative signals. To further refine the learned representation, a progressive residual mechanism is applied in Equation 12.where H t represents the feature representation at iteration t. This process of refinement helps the model to improve the quality of the features and the details of the differences between the mental health conditions. Lastly, the improved feature representation is used in the classification layer to determine the mental health state is given in Equation 13.where H f i n a l is the final refined multimodal feature representation obtained from the MPDSA-CD network, and y ^ denotes the predicted probability distribution over anxiety, depression, and mild cognitive impairment classes. The parameters of the network are optimized by the SFOA. The initial population is generated as in Equation 14.where X i j is the position of the i t h, l j and u j are the lower and upper bounds, and r ∈ 0 , 1 is a random number. During the exploration phase, solutions are given as Equation 15.where X i , p T and Y i , p T represent the current and updated positions, X b e s t , p T represents the best solution, a 1 and θ are used to control the search process. In the exploitation phase, the solutions are refined as given Equation 16.where denotes the distance between the best solution and a randomly chosen candidate, and r 1 , r 2 ∈ 0 , 1 are random coefficients used for updating. Thus, the integration of dense feature propagation, progressive self-attention, and SFOA-based optimization helps the MPDSA-CD model to effectively learn robust multimodal representations, resulting in accurate and reliable classification of mental health conditions. Besides classification, in this proposed framework, there is also a cognitive risk assessment module that will be used to quantify the severity level of cognitive decline and mental health conditions. The predicted class probabilities are transformed into a continuous risk score as given in Equation 17.where R is the cognitive risk score, P i represents the predicted probability of each mental health class, and w i refers to the severity weights for anxiety, depression, and MCI, which is considered an indicator of cognitive decline.

H l = σ W l F i * , H 1 , H 2 , . . . , H l − 1 + b l (8)

H d e n s e = H 1 , H 2 , . . . , H l (9)

Q = H d e n s e W Q , = H d e n s e W K , V = H d e n s e W V (10)

A = S o f t ⁡ max Q K T d (11)

H t + 1 = H a t t + H t (12)

y ^ = S o f t ⁡ max W c H f i n a l + b c (13)

X i j = l j + r u j − l j (14)

Y i , p T = X i , p T + a 1 X b e s t , p T − X i , p T cos θ , r ≤ 0.5 X i , p T + a 1 X b e s t , p T − X i , p T sin θ r > 0.5 (15)

Y i T = X i T = r 1 d m 1 + r 2 d m 2 (16)

R = ∑ i = 1 C w i · P i (17)

3.6 Model explainability using SHAP

To achieve interpretability and clinical transparency, SHapley Additive exPlanations (SHAP) is used to explain the predictions made by the MPDSA-CD model. SHAP provides a quantitative measure of how each feature, including various multimodal features (Choudhary et al., 2025) such as demographic, cognitive, behavioral, and physiological features, contributes to the final prediction and helps to identify important risk factors for anxiety, depression, and MCI. The prediction is given as a combination of feature contributions as given in Equation 18.where f x is the predicted output, ϕ 0 is the baseline prediction, and ϕ i represents the contribution of the i t h feature. Positive and negative SHAP values indicate the degree to which each feature increases or decreases the predicted risk. This approach offers individual-level as well as global-level insights, making it easier to increase model transparency, trustworthiness, and facilitate decision-making in a clinical context.

f x = ϕ 0 + ∑ i = 1 n ϕ i (18)

4 Result and discussions

The proposed multimodal mental health prediction framework was implemented using Python, which was then tested on a system with an NVIDIA GPU, e.g., NVIDIA RTX 3080, and an Intel i7 CPU, which can efficiently support the training process to achieve high computational performance, faster convergence, and accurate prediction of anxiety, depression, and mild cognitive impairment in real-time. Table 1 provides the hyperparameter values of the proposed method.

TABLE 1. Hyperparameter values of the proposed method.

4.1 Performance analysis of the proposed method

The performance of the proposed framework is examined using standard evaluation metrics in terms of accuracy, robustness, and efficiency in predicting anxiety, depression, and mild cognitive impairment.

Figures 3a,b the confusion matrix indicates that the most of the samples are correctly classified by MPDSA-CD and achieves near-near- classification accuracy for USA data samples (Depression 2064, Anxiety 1985 and Normal 1530) and India data samples (Depression 2064, Anxiety 1986 and Normal 1530) also reveals little misclassification on Anxiety samples. This focuses the strength, discrimination ability, and generalization performance of the model for predicting mental health.

The proposed MPDSA-CD framework performed better than the benchmark methods shown in Table 2. In this study the Kruskal-Wallis test was only used to select features and identify the discriminative multimodal attributes and is not meant to be interpreted as evidence of better classification performance. Typically repeated experimental runs, cross validation performance distributions and/or independent validation cohorts are necessary to formally test the competing predictive models. It was not possible to perform comparative hypothesis testing for such repeated performance distributions for all baseline methods, so this testing was not carried out in the present work. As a result, improvements in accuracy, precision, recall and F1-score provide some evidence for the effectiveness of the proposed framework. To validate the reported improvements in performance, future work will involve repeated k-fold cross-validation experiments, independent external validation cohorts and statistical comparison procedures like paired t-tests, Wilcoxon signed rank tests, McNemar’s test, and Friedman post hoc analysis.

TABLE 2. Performance comparison of the MPDSA-CD.

The proposed MPDSA-CD framework performed 99.8 % accuracy, precision 99.7%, recall 99.9% and F1 score 99.8 which is better than the benchmark methods shown in Table 2.

To quantify the relative improvement in performance offered by the proposed MPDSA-CD framework over representative state-of-the-art approaches, Table 3 quantifies this relative improvement. The proposed model constantly shows the improvements of all the evaluation metrics such as accuracy (from 1.79% to 4.39%), precision (from 2.09% to 4.31%), recall (from 1.99% to 4.78%) and F1-score (from 2.26% to 4.48%). It emphasized that the Kruskal–Wallis analysis applied in this work was only to select the features and to detect multimodal discriminative attributes of the data. Use of feature level statistical significance does not imply a better prediction ability. The formal assessment of predictive models typically requires multiple experimental runs, a distribution of performance measures from cross-validating the models or an independent validation cohort to substantiate hypothesis testing.

TABLE 3. Relative performance improvement of proposed model compared with existing methods.

Quantifies the relative improvement in performance offered by the proposed MPDSA-CD framework achieves 99.8 % accuracy which represents state-of-the-art approaches.

In the present study, comparative statistical significance testing was not done as all competing approaches did not have repeated performance distributions. This means that the gains that are reported taken in the context of what has been observed, not statistically proven to be superior. However, the steady improvement in all of the assessment metrics suggest the effectiveness and strength of the proposed MPDSA-CD framework. Future studies should involve repeated k-fold cross validation experiments, external validation cohort, and statistical comparison procedures (paired t-test, Wilcoxon signed rank test, McNemar’s test and Friedman post hoc analysis) to determine the statistical significance and reproducibility of the reported improvement in performance.

Figures 4a,b shows that the proposed model has achieved near-classification for the USA and India data sets with AUC = 1.00 for Normal, Anxiety, and depression classes. Such high efficiency is due to the effective multimodal feature integration by CDFLN, attention learning in MPDSA-CD, and hyperparameter tuning by means of the Starfish Optimization Algorithm, which improved discriminative representation learning, classification efficiency, and cross-population generalization performance.

Figures 5a,b indicates that the MPDSA-CD achieves performance for the USA and India samples because its AP score is 1.00. The system demonstrates accurate performance because it achieves precision and recall while maintaining high performance for distinguishing between different classes. The results confirm the model’s capability to accurately identify mental health conditions with minimal false positives and false negatives, thereby demonstrating strong robustness, prediction consistency and generalization across diverse multimodal population data.

Figures 6a,b highlight that the model achieves almost class metrics (precision, recall, specificity, F1-score), which reach values near 1.0 for the Normal, Anxiety, and Depression classes in both the USA and India patient samples. The balanced metric distribution demonstrates that the model identifies positive and negative cases with equal accuracy while it maintains consistent screening abilities across demographic groups.

Figure (a–c) Accuracy Curve: the accuracy curve shows the model’s performance over time during training. As training progresses, the accuracy steadily increases, indicating that the model is learning to correctly classify mental health conditions. Loss Curve Figure (b–d): The loss curve shows how the model’s error (loss) decreases over time during training. A decreasing loss indicates that the model is improving its predictions by minimizing errors. Figures 7a–d illustrates that the approach has obtained near-training and validation accuracy with minimal loss. This indicates this Hybrid model converges quickly and learns well without any overfitting.

Figures 8a,b shows that the correlation heatmap has been obtained after preprocessing for USA and Indian patient samples, in which the normalized and balanced features show strong correlations among the variables, including psychological, cognitive, and behavioral features, while the low correlation values indicate low redundancy, which will improve the quality of the data and make it more suitable for accurate model training and feature learning.

4.2 Feature learning analysis

This section evaluates the feature learning capability of the proposed MPDSA-CD framework by analyzing how effectively it extracts and represents multimodal information. The system evaluates three aspects of learned representations, which include reconstruction performance, feature correlations, and statistical significance, to highlight the representations effectiveness for mental health classification across different datasets.

Figures 9a,b shows that features that show significant results through high negative logarithm of p-value measurements serve as essential elements for differentiating between depression-related categories. The results show that these medically important characteristics increase feature distinctness while boosting prediction accuracy and enabling precise depression pattern identification across different patient groups.

Figures 10a,b indicates that the correlation of features varies in the USA and India data samples. Strong positive and negative values show redundancy and diversity in the features, which helps in the selection of features and the improvement of the classification of depression.

Figures 11a,b shows that the features with higher H-Statistics and lower p-values possess strong discriminative powers, with the top features being identified for both the USA and India samples, thus ensuring effective classification performance. The statistical results highlights that chosen multimodal indicators improve classification performance and model interpretability and maintain diagnostic accuracy across different population groups.

Figures 12a,b shows that all features in both the USA and India samples have p-values less than 0.05, indicating statistical significance. This further indicates that feature selection is effective for reliable classification performance. The results signify that the chosen feature selection method successfully maintains valuable multimodal attributes, resulting in improved prediction reliability, decreased duplicate data and enhanced mental health classification accuracy. Table 4 illustrates that 88.5% of the features reach statistical significance at the 0.05 alpha level because their p-values approach zero and their H-statistics produce high values, which show their ability to distinguish between different outcomes in model prediction.

TABLE 4. Kruskal-Wallis feature selection summary.

Figures 13a,b shows that the −log10 (p-values) distributions spread across the significant features of both the USA and India patients. The majority of features display statistical significance beyond the threshold of α = 0.05, which indicates their ability to differentiate between mental health groups. The selected multimodal attributes show that they improve feature reliability together with interpretability and classification performance.

Figures 14a–d illustrate that the reconstructed features closely match the original feature representations, which proves that the reconstruction mechanism successfully maintains vital multimodal data. The observed low reconstruction error across most feature indices shows strong representation learning ability, while the higher-error regions, which appear in specific areas, show minor information changes that do not impact overall classification accuracy or predictive performance.

Figures 15a,b shows the training and validation reconstruction loss curves, which display the performance of the feature extraction model on the USA and India datasets. The two losses show a gradual decrease until they reach stable values throughout the training periods, which focus successful learning and improvement of multimodal feature representations. The validation loss showed lower values than the training loss, which indicates that the system generalizes well while showing less overfitting and thus proving the feature extraction framework can deliver stable mental health prediction results.

Figures 16a–d illustrates that the majority of students fall within the no cognitive risk category, while only a smaller proportion, approximately 15%, are identified as being at cognitive risk. A similar distribution pattern is consistently observed across both the anxiety and depression groups, indicating stable risk stratification and demonstrating the model’s capability to effectively differentiate cognitive vulnerability levels among affected student populations.

4.3 Model interpretability analysis

This section examines the interpretability of the proposed MPDSA-CD framework using SHAP-based explanations to understand feature contributions toward predictions. It identifies key behavioral, psychological, and cognitive factors influencing model decisions, ensuring transparency. The analysis highlights how individual features positively or negatively impact mental health classification outcomes across datasets.

Figures 17a,b shows that three factors, which include memory loss and dietary habits, and emotional indicators, must be considered because they determine how the model makes its predictions. The SHAP analysis establishes a comprehensive interpretability system that measures the effect of each feature and shows how each element contributes to the prediction results with positive or negative effects. The results show how different behavioral patterns and psychological traits work together to impact mental health evaluation processes. The visualization process establishes model transparency, which builds trust in the system while showing how it makes decisions. The system produces mental health predictions that include risk evaluation results that are both accurate and understandable because of its data-driven approach.

SHAP analysis not only helps to find the influential predictors but also offers clinically relevant insights, which can guide early interventions and personalized mental healthcare. One of the most impactful features that were found by the proposed framework were memory loss indicators that showed strong contributions to the predictions of anxiety, depression and cognitive decline all across the board. Memories impairment is clinically important if it is elevated, as it can represent early neurocognitive dysfunction, chronic exposure to stress, decreased concentration, or the onset of mild cognitive impairment. People who show these patterns should be thoroughly evaluated with a neuropsychology assessment, should participate in cognitive rehabilitation and should be monitored on a regular basis. Other variables of diet influence were also found to be very powerful predictors. It has been shown in previous clinical studies that poor nutrition, higher psychological stress, emotional instability, sleep disturbances and depressive symptoms are associated with each other. Therefore, students with diet problems should be referred to diet counselor, lifestyle modification program and preventive mental health interventions to enhance their psychological health.

The behavioral measures, such as changes in daily routines, decreased social interest and changes in activity patterns, had a significant impact on the prediction results. This is typical behavior that is often seen as early symptoms of anxiety and depression. The proposed framework can thus help mental health workers and mental health counsellors to identify at-risk individuals before severe symptoms emerge so that psychological assistance and intervention can be provided at the right time. The emotional and psychological dimensions such as stress-related and mood-related measures showed high importance for prediction in both data sets. If the SHAP scores of these features are elevated, this may suggest that a targeted psychological counselling, a stress-management programme, a mindfulness-based intervention or a clinical psychiatric assessment may be helpful depending on the severity of the symptoms.

In addition, attention-related deficits, measures of thought disorganization and cognitive functioning were significant predictors of model predictions. All these factors could help the clinician distinguish between transient stress and ongoing cognitive deterioration which could guide more individualized treatment planning and risk stratification. The framework’s multimodal assessment approach can be complemented by SHAP to serve as an interpretable clinical decision-support system. The framework does not only make a diagnosis prediction but also identifies specific risk factors that correlated with an individual’s mental health status for more specific intervention strategies to be developed by healthcare professionals, psychologists and university counsellors based on patient-specific needs.

Figures 18a,b depicts the global feature importance based on mean SHAP values, thereby showing how much each particular feature affects the outcome of the model. From clinical point of view Diet is the topmost important factor, while memory loss and various other cognitive and behavioral factors like attention and thought disorder follow behind. Unhealthy eating habits patterns may correlate with increased stress and reduced emotional stability. Similarly memory loss score may result into mild cognitive impairment. These indicators may support the counselors, psychologist in practical decision and help to find at risk students and provide personalized mental health support, counseling, clinical treatment, monitoring of patient. making Some latent features that have been extracted also affect the output to some extent; however, their influence is comparatively lower. It becomes evident that lifestyle factors and variables related to cognitive impairment play a significant role in decision-making.

Figures 19a,b displays the two-stage MPDSA-CD decision system, which is used to achieve complete accuracy in diagnosing Depression and to assess cognitive risk at 89.5%, which designates the subject as “AT RISK”. SHAP analysis shows that three main factors, which include memory loss and behavioural changes and lifestyle indicators, help develop personalized clinical recommendations that doctors can use for treatment and ongoing patient assessment.

Evaluating a model through ablation studies helps understand the contribution of individual components by removing modules and observing performance degradation.

Table 5 indicate each module contributes to performance optimization; for instance, removing MPDSA-CD reduces accuracy to 96.8%, which highlights feature interacts with learning processes, while excluding SFOA lowers precision to 94.8%, which shows its value in achieving better parameter tuning and convergence results.

TABLE 5. Ablation study of the proposed method.

Indicates each module contributes to performance optimization, the whole architecture with all components achieves the accuracy 99.8%, precision 99.7%, and recall 99.9% which is overall enhance proposed model.

4.4 Data leakage prevention and generalization analysis

To avoid data leakage or overfitting, there were a number of measures taken to ensure the reliability of the reported results. The data set was split into separate training, testing and validation sets before training the models. Participant records did not appear in more than one subset in the partitioning process. The statistics computed in the preprocessing steps (normalization, feature selection, feature transformation) were only calculated from the training subset and then applied to the other two subsets, validation and test.

To minimize overfitting, dropout regularization, adaptive optimization, and monitoring through the use of validation sets were used during training. The improvement of training and validation performance in Figure 7 shows that there is no significant drop-off, suggesting that the algorithm learns steadily and has a good generalization ability. Moreover, the fact that the performance improvement is found to be due to the combination of Kruskal-Wallis feature selection, CDFLN feature extraction, MPDSA-CD classification and SFOA optimization instead of a single dominant component is confirmed by the ablation study presented in Table 4.

The proposed framework was able to attain the high predictive performance, but external validation by other cohorts was not conducted in the present study. Future studies will be directed towards multi-center validation and cross-institutional evaluation to further evaluate the model’s robustness, generalizability and clinical applicability in different populations.

4.5 Computational complexity, scalability, and deployment analysis

Besides predictive performance, the real-world implementation of AI-driven mental health assessment systems needs to take into account the computational efficiency, scalability, and feasibility of the operational system. The envisioned framework is comprised of three main components: feature extraction based on CDFLN, classification based on MPDSA-CD and optimization based on SFOA.

The computation of the CDFLN module is mainly determined by the modality specific feature learning and cross-directional feature interaction operations. The cross-modal interaction can be approximated to be O (M²d) where d is the dimension of features and M is the number of multimodal inputs that are computed pairwise between the modalities. However, the computational growth is still under control as the number of modalities is limited within the proposed framework. MPDSA-CD classifier uses the dense feature propagation and self-attention mechanisms. The dense connectivity helps in the feature reuse and gradient propagation among the layers, but the features are concatenated, which leads to the high memory usage. The self-attention mechanism adds computational costs on the order of O (n²) for the length of the sequence n or the size of the representation of the features. Self-attention, in spite of this extra cost, allows us to efficiently model long-range dependencies and relationships between domains that are hard to model with traditional recurrent architectures.

The Starfish Optimization Algorithm (SFOA) adds extra optimization overhead in training by sequentially searching over the population of parameters. At inference time, this cost is not applied after the model is deployed and so does not impact any inference time. As a result, during application, the optimized parameters enable efficient prediction and also optimization is enhanced with better parameter selection.

To implement the framework, it was written using Python, with NVIDIA GPUs used to accelerate the process. This is done offline, so that the training of costly components like feature extraction, attention learning, and optimization are done before deployment. In inference, only a forward-pass operation is needed, which allows for the prediction generation to be performed quickly for real-time or near real-time decision support applications. The architecture has the ability to scale when more multimodal inputs are added, and does not need to change the learning framework fundamentally. The modularity enables the incorporation of additional data sources such as signals from wearable sensors, neuroimaging data, electronic health records and conversational text from digital mental health assistants.

While the present work gives theoretical considerations of computational complexity and deployment feasibility, the focus of this work was not computational efficiency. Therefore, detailed statistics on training time, inference time, memory usage, size of the model and usage of hardware was not collected systematically. Although the proposed framework was successfully implemented on the GPU-accelerated computing resources, further research is needed to build a thorough benchmarking study that assesses the execution time, resource consumption, energy usage and deployment performance in cloud networks and edge computing networks. These analyses will help to gain another perspective in the practical scalability and applicability of the framework in the real-world.

The proposed framework is not proposed as a full autonomous diagnostic framework; it is proposed as a clinical decision support tool with regard to the feasibility of deployment. SHAP-based explainability will make the predictions more transparent and explainable, aiding the adoption of the model by healthcare practitioners, psychologists, and counsellors. Moreover, the framework could be used in cloud-based healthcare systems or institutional mental health monitoring systems that have the computational power to run the model. The results of the present study exhibit good prediction accuracy and potential to deploy this in practice, however, detailed benchmarking of training time, memory usage, inference time, and energy usage across various computing environments will be required to aid in the deployment of this to the real world on large scale.

The discussion has been revised to clarify that the reported performance reflects validation within the datasets available in this study and should not be interpreted as complete external clinical validation. The need for evaluation on independent institutional cohorts has been explicitly highlighted.

4.6 Discussion

The proposed MPDSA-CD approach proves increase in performance compared to the existing approaches on the USA and India patient samples, with an accuracy of 99.8%, surpassing the performance of DDNet (97.98%) and HYBRID-Bi-LSTM (95.41%). This approach offer next-generation AI framework that integrates Multimodal and offers a scalable, explainable, early detection and proactive solution to revolutionize student mental health support and health assistance based on Natural Interventions.

The confusion matrices prove the accuracy with minimal misclassification, while the ROC and precision-recall curves attain an AUC and AP score of 1.00, indicating discrimination. The selection of features proves that all p-values are less than 0.05, indicating statistical significance. The ablation test proves the performance of the approach by reducing the accuracy to 96.8% without the use of the MPDSA-CD approach. SHAP analysis highlights key features like memory loss and stress as major predictors.

This proposed hybrid approach introduces increasing computational complexity because of feature extraction, integration of multimodal inputs and deep learning based classification. This commences higher training cost compared to conventional approaches. Once the model is trained it can provide predictions efficiently for practical approach and early stage detection.

This Hybrid approach can be applied to large-size multimodal datasets. This can be used in Mental healthcare and Education sector for age group 18–30. The proposed framework has consistently achieved high performance across the India and USA datasets, but these tests were run as part of this study and were not a complete external test and do not imply complete validation. Thus, the reported performance needs to be understood as being well validated internally within geographically separate populations. The framework needs to be further validated with external healthcare institutions’ cohorts to confirm the robustness, reproducibility, and generalizability of the framework in real-world clinical settings. Going forward, multi-center studies will be conducted with a few diverse demographic and clinical populations.

The proposed framework showed excellent overall predictive performance across both India and USA cohorts, but the present study did not have a fully independent external validation dataset not collected into the development pipeline. Future studies will involve the assessment of the framework on external, independent, multi-institutional and geographically diverse cohorts to further assess the robustness, clinical reliability, and feasibility of deployment in real world.

The proposed architecture is compared with the conventional CNN-based, LSTM-based and transformer-based multimodal framework and is shown that it has three major advantages. Second, MPDSA-CD adopts dense feature propagation while incorporating self-attention mechanisms, which allows it to model the reuse of local features and the global context simultaneously. Third, SFOA is more efficient in optimization in complex multimodal feature spaces. All these mechanisms are complementary and contribute to better representation learning, robustness, and model interpretability, all of which are key properties of a reliable system for the assessment of mental health.

To demonstrate interpretation of the results of the experiments, all visualizations were modified to follow standard guidelines for scientific visualization. The sizes of font labels, legends and annotations for the axis were increased to make them easier to read both on screen and in print. Label space and layout were optimized for feature-based visualizations like SHAP summary plots, Kruskal-Wallis feature ranking plots and correlation matrices to reduce the number of overlapping labels. To identify the relationships of important features, higher resolution rendering and clearer placement of annotations were used to generate heat maps. The statistical feature rankings were re-designed and re-visualized to emphasize most informative, discriminative features, and reduce visual noise from less informative features. Similarly, SHAP visualizations were simplified to highlight the most relevant variables to the model’s prediction, and to the cognitive risk assessment. The confusion matrices and ROC, precision-recall curves and training history plots were also modified to use larger legends, better axis descriptions, and better color contrast to aid in interpretation. The changes increase the clarity, repeatability and presentation of the experimental analysis. Clarified in the discussion that the reported performance is due to validation within the datasets available in this study, and is not a complete external clinical validation. The importance of assessment of independent institutional cohorts has been specifically emphasized.

4.7 Analysis of the robustness and generalization

The proposed MPDSA-CD framework showed consistently good predictions for both India and USA cohorts for the adopted evaluation protocol. Ablation experiments, training and validation learning curves, model interpretation with SHAP and evaluation on two geographically distinct datasets were used as complementary analyses to assess model robustness. These analyses suggest that the proposed framework is able to learn stable multimodal representations and that it does not manifest clear symptoms of degradation under the conditions of these experiments.

However, in the present study a single hold-out validation strategy was used, and repeated experimental evaluations and confidence interval estimation and sensitivity analysis were not carried out. The results obtained therefore suggest good predictive empirical validity but do not provide conclusive proof of the generalizability across all individuals or statistical robustness. Future studies will continue to involve repeated stratified cross validation, construction of confidence intervals, sensitivity analysis with varied distributions and testing of the proposed framework in independent multi-centre external cohorts to further demonstrate the robustness, reproducibility and clinical applicability of the proposed framework.

5 Conclusion

The MPDSA-CD framework provides a highly efficient and interpretable DL solution for multimodal prediction of anxiety, depression, and MCI in young adults, along with cognitive risk assessment. This Hybrid approach demonstrates 99.8% accuracy over existing techniques. The CDFLN and MPDSA-CD facilitate efficient cross-domain feature learning, further optimized by the Starfish Optimization Algorithm. Discriminative multimodal indicators were identified through statistical validation at the feature level. Instead of direct statistical significance comparison between the proposed model and competing methods. We acknowledge that comparative statistical validation would boost the model and this is an important direction for future work. The proposed framework is also statistically robust and generalizable, with consistently high predictive performance on the available datasets, but further statistical robustness analyses (repeated cross-validation, confidence interval estimation, sensitivity analysis, and validation with independent external datasets) are needed to further validate the statistical robustness and generalizability of the proposed framework. The reported results are therefore considered as strong empirical evidence from the evaluation protocol used, and should not be interpreted as conclusive evidence for universal clinical use. One weakness of the present study is that no independent, external validation cohort was used. Future studies will examine the generalizability and applicability to real-world scenarios of the proposed framework using external validation sets from different institutions and populations. The proposed framework showed empirical predictive performance, but future studies will include formal statistical significance tests and independent validation experiments to provide statistical robustness and reproducibility of the reported improvement. Explicitly, it is now stated that another important avenue for future research is the need for external validation through independent cohorts from other healthcare institutions before a large-scale clinical deployment.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Rakesh Kumar Mahendran, Saveetha Engineering College, India

Reviewed by: Chaimae Taoussi, Université Sultan Moulay Slimane, Morocco

Kirti Seth, Inha University in Tashkent, Uzbekistan