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

Abstract

The increasing availability of wearable device data provides an opportunity for developing personalized models for health monitoring and condition prediction. Unlike conventional approaches that rely on pooled data from diverse individuals, our study explores the strategy of intentionally overfitting models to personal data and subsequently applying a transfer learning technique to refine performance for each user. We predicted Next-Day Condition (NDC) and Next-Day Emotion (NDC) while considering diverse features such as physical activity, sleep patterns, environmental context, and self-reported measures. Initial experiments showed that models trained at the sample level performed better on evaluation data but failed to generalize effectively during external validation. In contrast, our personalized learning approach, initiated with a pre-trained model, significantly enhanced accuracy within ten days of incremental user-specific training. Although generalization across the entire cohort diminished after individual tailoring, extended individualized training increased the overall predictive accuracy for each participant's personal data. The interpretation of feature importance using Shapley's additive explanations revealed substantial variability in the features influencing predictions across individuals, emphasizing the need for tailored health models. These findings highlight the potential of combining intentional overfitting and transfer learning in constructing high-performance user-specific predictive models from wearable data. Future research should expand the number of participants, extend the training period, and refine these methods to bolster personalized digital health solutions.