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

TL;DR

A GAN-based system improves early detection of infectious diseases by imputing missing wearable sensor data, even in low-resource settings.

Contribution

A lightweight GAN framework that imputes missing heart rate data and enables early infection detection using rule-based anomaly detection.

Findings

The system triggered early alerts in 100 malaria cases, including 42 solely through imputation.

Alerts occurred 11.9 days before symptom onset, matching the parasitemia window.

The GAN reduced reconstruction error by 58% when generalizing from COVID-19 data to malaria.

Abstract

High rates of missing data in wearable sensor streams hinder early detection of infectious diseases, especially in low-resource settings with inconsistent device adherence and connectivity. We developed a lightweight generative adversarial network (GAN) framework that imputes missing heart rate data and integrates with a rule-based anomaly detection algorithm to identify early signs of infection. In a cohort from rural Kenya (n = 300, 161 malaria-positives), our system triggered early alerts in 100 cases, including 42 solely with imputation. Alerts preceded symptom onset by 11.9 days, aligning with the 11.7-day parasitemia window from controlled trials. Despite 50% data coverage, alerts occurred on 3.5 consecutive days during the infection window, improving early detection by 35%. The GAN, trained only on external COVID-19 data (n = 3318), generalized to malaria, reducing reconstruction error by 58%. This approach demonstrates scalable, cross-pathogen physiological monitoring and offers a robust tool for disease surveillance in settings challenged by high wearable data loss.

Figures