Explainable AI Reads Memory, Sleep and Diet to Flag Student Depression Early

Anxiety, depression and mild cognitive impairment are quietly eroding the academic lives of young adults between 18 and 30, and clinicians have long struggled to catch these conditions early because diagnosis still leans on subjective questionnaires and limited clinical interviews. Now, researchers at Amity University in Mumbai have built an artificial intelligence framework that reads a person’s mental state from 53 multimodal indicators spanning demographics, cognition, behavior, physiology and psychology, then explains exactly why it reached its conclusion. The system, described in Frontiers in Signal Processing, achieved 99.8 percent accuracy, 99.7 percent precision, 99.9 percent recall and an F1-score of 99.8 percent across two geographically separate datasets of 27,901 participant records each, drawn from clinical settings in India and the United States.

The framework, developed by Rupali D. Kasar and Garima Shukla, addresses a persistent weakness in AI-based mental health prediction: most existing models rely on single-modal or weakly fused data, such as Twitter text alone or survey responses alone, and their deep learning cores operate as opaque black boxes that clinicians cannot audit. The new pipeline begins with rigorous preprocessing. Missing numerical values are imputed using mean, median or K-Nearest Neighbor strategies depending on the complexity of the feature dependencies, categorical variables are encoded with label or one-hot schemes to avoid imposing false order, and textual survey data undergoes tokenization, lowercasing and character removal before being converted into numerical form. Min-Max normalization then squeezes every numerical feature into the range [0,1], preventing variables with large raw values from dominating the learning process.

Before any deep learning begins, a Kruskal-Wallis test acts as a statistical gatekeeper. This non-parametric test checks whether the distribution of each feature differs significantly across the anxiety, depression and MCI groups, preserving only features whose p-values fall below the 0.05 significance level. Of the 53 multimodal parameters tested, 48 — roughly 88.5 percent — proved statistically significant, with H-statistics ranging from 17.53 to 27,894.84. The single most discriminative feature turned out to be the disorganized thought expression score, a finding that aligns with clinical understanding of how thought disorganization marks emerging cognitive dysfunction. This selection step reduces dimensionality, cuts redundancy and ensures that only genuinely informative behavioral and cognitive factors feed the predictive models.

The heart of the framework is the Cross-Directional Feature Learning Network, or CDFLN, which learns modality-specific representations through non-linear transformations while simultaneously modeling pairwise interactions between modalities. Each modality branch passes through an Adaptive Layer Normalization Unit to stabilize feature distributions and reduce modality-specific bias, and cross-modal interaction weights are learned through element-wise multiplication between modality pairs. The final representation aggregates both intra-modal characteristics and inter-modal dependencies, allowing the network to discover latent relationships — such as the combination of cognitive decline, emotional instability and behavioral change — that simple feature concatenation would miss. This matters because mental health conditions are multifaceted products of interconnected factors, not isolated variables.

Classification is handled by the Multi-model Progressive Dense Self-Attention for Cross Domain architecture, MPDSA-CD. Dense feature propagation means every layer receives the concatenated output of all preceding layers, maximizing feature reuse and gradient flow while minimizing information loss as signals travel through the network. Self-attention then projects features into query, key and value spaces, computing normalized attention weights that dynamically highlight clinically relevant indicators and suppress redundant signals. A progressive residual mechanism refines the representation across iterations, sharpening the subtle differences between anxiety, depression and MCI before the final classification layer produces probability distributions over the three conditions. The architecture also explicitly models cross-domain relationships, jointly learning behavioral-cognitive and physiological-demographic dependencies that are crucial for identifying subtle mental health patterns.

Hyperparameter tuning and parameter refinement are entrusted to the Starfish Optimization Algorithm, a bio-inspired metaheuristic that generates an initial population of candidate solutions within defined bounds, then alternates between exploration and exploitation phases to search the parameter space. During exploration, solutions are updated using controlled coefficients and the best solution found so far; during exploitation, refinement uses the distance between the best solution and randomly chosen candidates. SFOA reduces convergence instability and the probability of getting trapped in local optima — a real risk when multimodal feature interactions produce highly non-convex objective surfaces. Ablation studies confirm its value: removing SFOA drops precision from 99.7 to 94.8 percent, while removing the MPDSA-CD module itself lowers accuracy to 96.8 percent, demonstrating that no single component carries the performance alone.

To make the system clinically transparent, the researchers applied SHapley Additive exPlanations, which decomposes every prediction into the contribution of each feature relative to a baseline. Positive SHAP values indicate features that increase predicted risk; negative values indicate protective factors. The global feature importance ranking placed diet at the top, followed by memory loss, attention deficits, thought disorganization and various behavioral and emotional indicators. Elevated memory loss scores may signal early neurocognitive dysfunction, chronic stress exposure or the onset of mild cognitive impairment, warranting neuropsychological assessment and regular monitoring. Poor dietary patterns correlate with higher psychological stress, emotional instability, sleep disturbance and depressive symptoms in prior clinical literature, suggesting students flagged for diet problems could benefit from nutritional counseling and lifestyle modification programs. Behavioral measures such as disrupted daily routines and reduced social interest — classic early symptoms of anxiety and depression — also carried significant predictive weight.

Beyond classification, the framework includes a cognitive risk assessment module that converts predicted class probabilities into a continuous risk score using severity weights for anxiety, depression and MCI. In the evaluated populations, roughly 15 percent of students were identified as being at cognitive risk, with similar distribution patterns across both anxiety and depression groups, indicating stable risk stratification. The two-stage decision system achieved complete accuracy in diagnosing depression and assessed cognitive risk at 89.5 percent, designating flagged subjects as AT RISK. The researchers emphasize that the framework is intended as a clinical decision-support tool rather than an autonomous diagnostician, with SHAP-based explanations helping psychologists, counselors and healthcare practitioners understand and trust the predictions before acting on them.

The evaluation was conducted on an NVIDIA RTX 3080 GPU with an Intel i7 CPU, using a batch size of 32, 100 epochs, a learning rate of 0.001, a dropout rate of 0.4 and eight attention heads. Confusion matrices showed minimal misclassification across USA samples (2,064 depression, 1,985 anxiety, 1,530 normal) and India samples (2,064 depression, 1,986 anxiety, 1,530 normal), while ROC and precision-recall curves reached AUC and average precision scores of 1.00 for all three classes in both cohorts. Compared with benchmark methods, the framework improved accuracy by gains ranging from 1.79 percent over Hy-OCNN-BiLSTM to 4.39 percent over HYBRID-Bi-LSTM, with corresponding precision, recall and F1 improvements. Training and validation losses converged smoothly with validation loss tracking below training loss, indicating good generalization without overfitting.

The authors are candid about limitations. Data leakage was prevented by splitting participant records into separate training, validation and test sets, with preprocessing statistics computed only from the training subset, yet no independent external validation cohort was used, and comparative statistical significance testing between the proposed model and baselines was not performed because repeated performance distributions were unavailable. Reported gains therefore represent strong empirical evidence within the evaluation protocol rather than conclusive proof of universal clinical superiority. Future work will involve repeated k-fold cross-validation, confidence interval estimation, sensitivity analysis, paired t-tests, Wilcoxon signed-rank tests, McNemar’s test and Friedman post hoc analysis, alongside multi-center validation across diverse demographic and clinical populations. The modular architecture is designed to scale, with the cross-modal interaction cost approximated at O(M²d) for M modalities and self-attention at O(n²), while SFOA’s optimization overhead applies only during training and not at inference, enabling real-time or near-real-time decision support once the model is deployed.

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

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

Article References: Kasar, R. D., & Shukla, G. (2026). An explainable AI framework integrating deep learning and large language model for student’s mental health. Frontiers in Signal Processing, 6, Article 1893420. https://doi.org/10.3389/frsip.2026.1893420

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1893420

Keywords: explainable, framework, integrating, deep, learning, large, language, model, student, mental, health, scientific research