From smartphone data to clinically relevant predictions: A systematic review of digital phenotyping methods in depression

Mobility-Based Smartphone Digital Phenotypes for Unobtrusively Capturing Everyday Cognition, Mood, and Community Life-Space in Older Adults: Feasibility, Acceptability, and Preliminary Validity Study (Preprint)

“…Specifically, they travelled farther, spent less time at home, and had greater diversity in the locations they visited. These findings are consistent with our hypotheses and with prior studies showing greater physical activity, less time at home, greater life space (the extent of movement through the environment during daily functioning), and engagement in a variety of activities (environmental complexity) are associated with better cognition, less depression, and reduced risk for MCI and dementia/ADRD [63,100,101,[116][117][118][119][120][121][122][123][124].…”

“…We examined many individual and contextual influences on mobility features to inform future studies in selecting covariates or moderating factors. This is a critical open question in the field of digital phenotyping [63], and our preliminary results provide important insights about which sociodemographic, contextual, and technology factors should be considered when interpreting this data. Age, education, sex, race, and cohabitation status appeared to be minimally associated with most mean-level and variability metrics, providing partial support for the objectivity of digital phenotyping features.…”

“…With regard to our technical protocol, use of an open-source platform and a publicly available data processing script facilitates replicability which is key to ongoing validation efforts [151]. Attempts at minimizing missing data using an automated checking script represent another strength and should be incorporated in future studies given the impact of missing data on subsequent findings [63]. Finally, our device-agnostic approach leveraging personallyowned smartphones versus a study issued device represents a notable strength as it promotes the naturalistic, unobtrusive nature of this method and affords increased scalability.…”

“…It collects high-frequency, finegrained data reflecting everyday behaviors "in the wild" without relying on user engagement, subjective report, or burdensome procedures. Preliminary support has been demonstrated in psychiatry studies leveraging a host of sensors (e.g., Global Positioning System [GPS], accelerometer, WiFi/Bluetooth signals, ambient sound and light, application use, call and text message metadata, keystroke dynamics) and imputed behavioral features (e.g., time spent at home, sleep cycles, level of socialization, routine/anomalies) to predict clinical outcomes including depression and bipolar disorder symptoms, suicide risk, psychosis relapse, and depression treatment response [58][59][60][61][62][63][64][65][66][67]. In the context of neurodegenerative disorders, several sensors -particularly keystroke dynamics and phone/battery use metrics -have shown associations with neuropsychological test performance, diagnostic severity, and even gray matter volume in clinical cohorts including those with MCI, AD, frontotemporal dementia, and multiple sclerosis [68][69][70][71][72][73][74].…”

“…Further, many existing studies have used exploratory approaches without a priori hypotheses [78]. As described by Hackett and Giovannetti (2022) and Leaning et al (2024), some of the many interpretive and logistical challenges of digital phenotyping can be mitigated with conceptual models and clinically-informed features to provide context to results and improve reproducibility [7,63].…”