Longitudinal Health Foundation Model: A Multimodal Self-Supervised Framework for Behavioral Health Prediction

This open-source project builds a self-supervised multimodal foundation model that integrates three major data sources—wearable devices, smartphones, and climate data—specifically for longitudinal behavioral health prediction. The model features advanced capabilities such as fairness auditing, climate generalization, and interpretability analysis, aiming to address core challenges in the digital health field, including the integration of multi-source heterogeneous data and long-term health status tracking and prediction.

With the popularization of wearable devices and smartphones, health data acquisition has become convenient. However, effectively integrating heterogeneous data sources and building AI systems for long-term tracking and prediction remain core challenges in the digital health field. Traditional health prediction models often rely on a single data source, lack in-depth modeling of the time dimension, and struggle to handle complex interactions of multimodal data.

Multimodal Data Fusion

The model integrates three types of heterogeneous data sources: wearable devices (physiological indicators such as heart rate, steps, and sleep), smartphones (behavioral indicators such as screen usage, location, and social interactions), and climate data (environmental factors such as temperature, humidity, and air quality), comprehensively reflecting the multi-dimensional impacts on physical and mental health.

Self-Supervised Learning Architecture

Adopting a self-supervised learning paradigm, it does not require large amounts of labeled data. Through tasks such as predicting future states and reconstructing masked segments, it learns dynamic health patterns from massive unlabeled time-series data.

Longitudinal Time Modeling

Designed for longitudinal data, it can capture long-term trends and seasonal changes in health indicators, model individual health evolution trajectories, and predict future health risks.

Fairness Auditing Mechanism

It has a built-in comprehensive fairness auditing function that detects differences in prediction results across different population subsets (age, gender, region, health status, socioeconomic background) to ensure non-discriminatory predictions.

Climate Generalization Capability

It adapts to data distribution differences across different climate regions, learns the transfer rules of climate-health correlations, and supports globally usable health prediction systems.

Interpretability Analysis

Using integrated gradient methods, it provides feature importance attribution, time importance analysis, and modal contribution decomposition to help understand model decisions and enhance user trust.

The model demonstrates value in multiple scenarios:

• Mental Health Monitoring: Identify early signs of depression, anxiety, etc., to support timely intervention;
• Chronic Disease Management: Track disease trends such as diabetes and cardiovascular diseases, and predict the risk of acute events;
• Health Behavior Intervention: Recommend targeted lifestyle adjustment suggestions;
• Public Health Research: Help understand the relationship between environment, digital lifestyle, and population health to support policy-making.

The project provides important insights for the health AI field:

1. Data Diversity: Fusion of multi-source heterogeneous data builds a more comprehensive health profile, which is better than a single data source;
2. Potential of Self-Supervised Learning: Under the scarcity of medical labeled data, it effectively uses unlabeled data and reduces reliance on labeling;
3. Necessity of Responsible AI: Treating fairness and interpretability as core functions meets the ethical requirements of medical AI.

Future development directions include:

• Real-Time Health Monitoring: Evolve from periodic assessment to continuous real-time monitoring;
• Personalized Medical Integration: Integrate molecular data such as genomics and proteomics;
• Digital Therapy Support: Provide intelligent assessment and feedback for prescription digital therapies. This project promotes the evolution of health AI towards comprehensive health understanding and is expected to realize more personalized, preventive, and inclusive medical services.