Introduction
Mental health challenges among college students have reached a critical point, with nearly 60% of students reporting overwhelming anxiety. While mobile sensing—using smartphone data to track activity, sleep, and social patterns—offers a continuous window into student well-being, most existing machine learning models operate as "black boxes." These models may predict a mental health crisis but often fail to explain the underlying behavioral drivers or offer personalized insights necessary for clinical intervention.
To address these limitations, a new framework called I-HOPE (Interpretable Hierarchical mOdel for Personalized mEntal health prediction) was introduced at CHASE '25. I-HOPE bridges the gap between raw sensor data and human-understandable behavioral categories, achieving high predictive accuracy while maintaining interpretability.
The I-HOPE Framework
The core innovation of I-HOPE is its two-stage hierarchical structure, designed to transform noisy sensor data into actionable insights.
Stage 1: Raw Features to Interaction Labels
Instead of feeding raw data (such as GPS coordinates or screen unlock frequency) directly into a classifier, I-HOPE maps 35 engineered features into five "Interaction Labels": Leisure, Me Time, Phone Time, Sleep, and Social Time. This stage utilizes a rule-based scoring system refined by Random Forest feature importance to ensure the labels are both accurate and meaningful to human observers.
Stage 2: Interaction Labels to PHQ-4 Category
These five behavioral scores are then processed by a Neural Network (Multi-Layer Perceptron) to predict the student's mental health status. Predictions are categorized based on the PHQ-4 scale: Normal, Mild, Moderate, or Severe. Crucially, the model is personalized; it is trained on a per-user basis, allowing it to learn that specific behaviors (like phone usage at home) may have different psychological implications for different individuals.
Dataset and Population Characteristics
The study utilizes the College Experience Study (CES) dataset, released by Dartmouth College in 2024. This represents the longest longitudinal mobile sensing dataset currently available, spanning five years (2017–2022). It captures the behaviors of 217 students across pre-pandemic, mid-pandemic, and post-pandemic periods.
The dataset provides a unique opportunity to observe how external stressors shift behavioral patterns. As shown in the distribution of PHQ-4 scores, while the majority of data points fall into the "Normal" category, there is significant representation of "Mild," "Moderate," and "Severe" statuses, providing the necessary variance for robust model training.
Key Insights
The researchers identified several critical factors in effective mental health prediction:
- Contextual Significance: Phone usage is not inherently negative. For instance, increased phone use in social or dining settings often correlated with poorer mental health (potentially indicating social anxiety), whereas phone use at home often suggested positive social connection.
- The Hierarchy Advantage: By introducing an intermediate interpretable layer, the authors actually improved accuracy. Moving from raw features (60-70% accuracy) to the I-HOPE hierarchy (91% accuracy) suggests that behavioral abstraction helps filter out environmental noise.
- Personalization: Feature analysis revealed that while "Duration of Walking" was a top predictor for 90% of students' leisure time, the remaining 10% relied more heavily on social interactions or phone calls. A generalized model would likely overlook these individual differences.
Results and Performance
I-HOPE achieved an overall accuracy of 91% in predicting PHQ-4 categories, significantly outperforming several baseline models:
- Aggregated Global Model: 60% accuracy.
- Personalized Model (All Features): 70% accuracy.
- Personalized Model (Top 50% Features): 65% accuracy.
Beyond accuracy, the model's "glass-box" nature allows for the generation of individualized behavioral profiles, enabling researchers and clinicians to see exactly which habits are the strongest predictors for a specific student's mental state.
Conclusion and Future Directions
I-HOPE represents a significant step toward making mobile sensing a practical tool for mental health support. By turning "black-box" predictions into an interpretable hierarchy, it offers a blueprint for AI that not only detects issues but helps us understand them. Future work will focus on increasing data granularity (moving from daily to hourly analysis) and improving data imputation methods to handle missing sensor values more robustly.
Summary updated