Fusion Strategy Evaluation for Clustering Depression Subtypes Using Multimodal Physiological and Social Data.

Area of Science:

• Digital health
• Computational psychiatry
• Machine learning for mental health

Background:

• Wearable and phone sensors offer potential for depression monitoring.
• Integrating diverse sensor data for mental health remains a challenge.
• Multimodal embeddings can represent complex temporal and inter-modal data patterns.

Purpose of the Study:

• To evaluate fusion strategies for multimodal embeddings in depression symptom clustering.
• To assess the impact of embedding space on clustering effectiveness.
• To identify optimal methods for integrating physiological and social sensor data.

Main Methods:

• Longitudinal dataset with physiological (ECG, accelerometer, respiration) and social (mobility, Bluetooth) data.
• Embedding generation using long short-term memory (LSTM) autoencoders.
• Clustering using K-Means and Gaussian Mixture Models (GMM).
• Evaluation using Beck Depression Inventory-II (BDI-II) scores and a custom BDI-Variance Ratio Clustering Score (BDI-VRCS).

Main Results:

• Early fusion with LSTM and GMM achieved the highest BDI-VRCS (0.3309).
• Early fusion outperformed mid and late fusion strategies (0.112 and 0.132).
• Social features were crucial for capturing depressive symptoms.

Conclusions:

• Early multimodal data integration is valuable for depression monitoring.
• This approach supports holistic, data-driven tools for early detection.
• Findings pave the way for personalized mental health interventions.