Towards Transparent Mental Health Insights: An Explainable AI Model for Career-Related Depression and Anxiety Among University Students Using Structured Data

Technical Summary: Towards Transparent Mental Health Insights

Problem Statement
The paper addresses the rising incidence of career-related anxiety and depression among university students, a challenge exacerbated by limited access to timely mental healthcare, the stigma surrounding help-seeking, and the resource-intensive nature of conventional diagnostic practices. Traditional methods rely heavily on subjective clinical observations of verbal and non-verbal cues, which are often inaccessible in low-resource environments. Furthermore, existing machine learning approaches frequently suffer from a lack of cultural sensitivity (often focusing on Western populations), poor integration of multimodal data, and a "black box" nature that hinders clinician trust. The authors identify a specific gap in detecting career-related mental health issues within diverse educational systems, such as those in Pakistan, while maintaining data privacy and model interpretability.

Methodology
The proposed framework is an Explainable AI (XAI) system designed to detect career-related depression and anxiety using structured behavioral data and facial emotion features. The methodology is built upon the following components:

• Data Source: The study utilizes the "Student Mental Health Survey" dataset, comprising structured responses (age, gender, GPA, academic satisfaction, financial concerns, etc.) from 512 university students in Pakistan. The dataset is labeled into "Depressed" (moderate to severe) and "Non-Depressed" (mild or no symptoms).
• Multimodal Feature Extraction:
  • Structured Data: Demographic and situational attributes.
  • Facial Features: Using the OpenFace toolkit, the system extracts head posture (pitch, yaw, roll), gaze direction (radian angles), and Facial Action Units (AUs) including presence and intensity.
• Federated Learning (FL): To preserve privacy and enable collaboration across institutions without sharing raw data, a Federated Learning paradigm is employed. Local models are trained on institutional data, and only model updates (gradients/weights) are aggregated centrally.
• Model Architecture: The core of the system is a Multimodal Neural Network (MNN) utilizing an Intermediate Fusion strategy. This approach aggregates outputs from various modalities before the final prediction layer, rather than fusing raw inputs (early fusion) or decision outputs (late fusion).
  • Three architectures were evaluated: Bi-LSTM, Window Block LSTM (WBLSTM), and Transformer.
  • Attention Mechanisms are integrated to prioritize significant features from different data sources.
• Label Smoothing: To address the inherent subjectivity and uncertainty in labeling mental health states (particularly distinguishing mild/moderate depression), label smoothing is applied. This technique softens the hard labels during training to improve generalizability and reduce overfitting.
• Explainability (XAI): The system incorporates Integrated Gradients (IG) and SHAP (SHapley Additive exPlanations) to trace predictions back to input features, ensuring transparency for mental health practitioners.

Key Results
Experimental evaluation was conducted on the Student Mental Health Survey dataset, comparing various fusion strategies and model architectures:

• Performance Metrics: The Window Block LSTM (WBLSTM) with Intermediate Fusion and Label Smoothing (0.05, 0.95) achieved the best overall performance:
  • Accuracy: 91.67%
  • Macro F1-Score: 88.89%
  • Precision: 91.40%
  • Recall: 87.03%
  • Micro F1-Score: 91.44%
• Comparison with Baselines: The proposed models significantly outperformed baseline architectures (standard Bi-LSTM and Transformer without advanced fusion or smoothing). Baseline models struggled with class imbalance, achieving accuracies around 68.75% and failing to identify true positives effectively (e.g., baseline Bi-LSTM had 0 True Positives).
• Label Smoothing Impact: The application of label smoothing was critical. For the Bi-LSTM model, a (0.3, 0.7) smoothing configuration improved accuracy to 91.67% and eliminated false positives, demonstrating the technique's ability to handle noisy labels and improve robustness.
• Feature Attribution: Integrated Gradients analysis revealed that specific behavioral markers were most influential in predicting depression. Key positive contributors included AU07_c (lip corner depressor), AU26_c (jaw drop), and AU25_c (lips part), as well as Gaze_y and Pose_Ry (head rotation). These findings align with psychological theories regarding social withdrawal and reduced facial expressiveness in depressed individuals.

Significance and Claims
The paper claims that its primary contribution lies in the development of a scalable, privacy-preserving, and culturally responsive AI system for mental health pre-diagnosis. By combining Federated Learning with Multimodal Intermediate Fusion, the study demonstrates that high-accuracy detection is possible without compromising student privacy or data sovereignty.

The authors emphasize the importance of Explainability in this domain. By using IG and SHAP, the model moves beyond a "black box," identifying specific, interpretable behavioral markers (such as avoidance of direct gaze and lower facial expressiveness) that align with psychological theory. This transparency is presented as essential for building trust among mental health professionals and facilitating the integration of AI tools into student support services.

The study concludes that intermediate fusion architectures, particularly when paired with label smoothing, offer a robust solution for handling the nuances and class imbalances inherent in mental health data. The authors position this work as a step toward creating ethical, interpretable, and effective AI solutions that can be adapted for global student populations, while acknowledging the need for future research to expand datasets to more diverse, global populations and incorporate real-time intervention capabilities.