Personalized Health Prediction AI Models Using Transfer Learning and Strategic Overfitting on Wearable Device Data

Area of Science:

• Digital Health
• Machine Learning in Healthcare
• Wearable Technology

Background:

• Wearable devices generate vast amounts of personal health data.
• Traditional health models use pooled data, limiting personalization.
• Developing user-specific predictive models is a key challenge.

Purpose of the Study:

• To investigate a personalized learning approach for health monitoring using wearable data.
• To evaluate the effectiveness of intentional overfitting combined with transfer learning.
• To predict Next-Day Condition and Next-Day Emotion using user-specific models.

Main Methods:

• Intentionally overfitting machine learning models to individual user data.
• Applying transfer learning to refine personalized models.
• Utilizing features from physical activity, sleep, environment, and self-reports.
• Employing Shapley's additive explanations for feature importance analysis.

Main Results:

• Personalized models significantly enhanced accuracy within ten days of user-specific training.
• Intentional overfitting and transfer learning improved individual predictive performance.
• Feature importance varied substantially across individuals, underscoring the need for tailored models.
• Generalization across the cohort decreased post-individual tailoring, but personal data accuracy increased.

Conclusions:

• Combining intentional overfitting and transfer learning shows promise for high-performance, user-specific predictive models.
• Personalized digital health solutions can be bolstered by these advanced machine learning techniques.
• Future research should focus on larger cohorts and extended training periods for further refinement.