Oral
in
Workshop: Workshop on Learning from Time Series for Health
Forecasting Exercise Lapses in Individuals with Type 1 Diabetes Using State Space Models
Rajat Doshi · Yunyang Li · Garrett Ash · Jason Liu · Laura Nally · Matthew Stults-Kolehmainen · Stuart Weinzimer · Lisa Fucito · Mark Gerstein
Keywords: [ sensors ] [ state space model ] [ mobile phone ] [ behavior ] [ diabetes ] [ time series transformers ]
Exercise is important for those with type 1 diabetes (T1D), but T1D poses exercise barriers including blood glucose destabilization and fear of hypoglycemia. To investigate these barriers, we measured the day-to-day variation of the barriers and used them to predict the occurrence of days without exercise (i.e., ”exercise lapses”). The study participants were 17 adults with T1D without regular exercise routines. They wore biosensors and completed real-time surveys to track exercise, mood, and sleep during 10 weeks of a flexibly-timed, beginner-level home exercise program. We leverage various machine learning techniques, consisting of logistic regression, random forest, time series transformers, and Mamba, a state-of-art state space model (SSM), for forecasting exercise lapses. We demonstrate that we can achieve 75.55 ± 2.6% accuracy with SSM, an improvement over the top baseline accuracy of 72.06 ± 2.9% achieved by classical ML techniques.