Overcoming Data Loss in Wearable Disease Detection with GAN-Based Imputation.

Area of Science:

• Digital Health
• Infectious Disease Surveillance
• Machine Learning

Background:

• High rates of missing data in wearable sensor streams impede early detection of infectious diseases.
• Low-resource settings face challenges with device adherence and connectivity, exacerbating data loss.
• Existing methods struggle to reliably impute physiological data for timely disease surveillance.

Purpose of the Study:

• To develop and evaluate a lightweight generative adversarial network (GAN) framework for imputing missing heart rate data.
• To integrate the GAN imputation with an anomaly detection algorithm for early identification of infections.
• To assess the system's effectiveness in a real-world setting with significant data loss.

Main Methods:

• Developed a lightweight generative adversarial network (GAN) framework for heart rate data imputation.
• Integrated the GAN with a rule-based anomaly detection algorithm for infection alerts.
• Validated the system on a cohort in rural Kenya (n=300) with high rates of missing data.

Main Results:

• The system triggered early alerts in 100 cases, with 42 alerts solely due to imputation.
• Alerts preceded symptom onset by an average of 11.9 days.
• Despite 50% data coverage, early detection improved by 35%, with alerts occurring during the infection window.
• The GAN, trained on COVID-19 data, generalized to malaria, reducing reconstruction error by 58%.

Conclusions:

• The developed GAN framework offers a scalable solution for physiological monitoring despite high wearable data loss.
• This approach enhances early infectious disease detection and provides a robust tool for disease surveillance in challenging environments.
• The cross-pathogen generalization capability of the GAN highlights its potential for broad public health applications.