Digital Phenotyping and Mental Health: Evaluating the Reliability of Passive Smartphone Behavioral Features Using a Public Beiwe Dataset

Abstract

Background/Objective: Digital phenotyping is the use of data collected from personal digital devices to study patterns in human behavior over time. It has become useful in mental-health research because smartphones can collect information in everyday settings. The objective of this research is to examine how reliable several commonly used smartphone-based behavioral features are. Methods: I used a publicly available dataset from the Beiwe digital phenotyping platform. The data included six participants and 49 participant-days from March 2022. I added Gaussian noise to the GPS coordinates at eight levels, with standard deviations of 0-800 meters in each direction. Results: Added GPS noise greatly increased total distance and mean speed. Time at home fell to 39% of baseline at 100 meters of noise and 1% at 800 meters. Radius of gyration remained within 1-7% of baseline through 200 meters. Changes in time at home and radius of gyration matched the values calculated from the amount of noise. Location entropy and number of significant places were less predictable because both depend on clustering. For 33 participant-days, GPS-based and self-reported time at home had a weak correlation in hours, which increased when expressed as proportions. However, day-to-day agreement within individuals was negligible. Conclusions: Behavioral features can respond very differently to the same sensor error. This study does not test psychiatric diagnoses. Instead, it checks how reliable smartphone measurements are before they are used in mental-health research. Keywords: digital phenotyping, mental health, smartphone sensing, Beiwe, GPS, mobility, behavioral features, privacy, measurement reliability, machine learning