Observation Time vs. Performance in Digital Phenotyping

Abstract

Mobile health (mHealth) technologies enable frequent sampling of physiological and psychological signals over time. In our recent work we used a convolutional neural network (CNN) model to predict self-reported phenotypes of chronic conditions from step and sleep data recorded from passive trackers in free living conditions. We investigated the impact of the time-granularity of the collected data and showed that training the models on higher- resolution (minute-level) data improved classification performance on conditions related to mental health and nervous system disorders, as compared to using only day-level totals. In the present work we shift the focus from the time resolution of the observation window to its duration. We study how the performance of the best-performing model on the highest-resolution data changes as the length of the data collection window is varied from 3 to 147 days for each user. We found that for mental health and nervous system disorders, a model trained on 30 days of mHealth data attains the same performance as using the full 147-day window of data, in terms of AUC increase over a baseline model that uses only demographics, height, and weight. Additionally, for the same cluster of conditions, only 7 days of data are sufficient to realize 62% of the maximum increase in AUC over baseline attainable using the full window. The results suggest that for some conditions health-related digital phenotyping in free-living conditions can potentially be performed in a relatively short amount of time, imposing minimal disruptions on user habits.