Background: Digital phenotyping involves moment-to-moment quantification of the individual experience using data from personal digital devices such as Fitbits and smartphones. These methods offer promise for prediction and early detection of clinical deterioration, but thoughtful approaches to missing data are essential for validity. We aimed to explore the relationship between clinical symptoms and compliance with digital phenotyping protocols to better understand missing considerations.
Methods: We used data from two digital phenotyping studies: a 9-month study of patients with bipolar disorder (BD; N=39) involving Fitbit monitoring and bi-weekly digital self-reports and a 3-month study of women with episodic migraine (N=30) involving Fitbit monitoring and both daily and weekly digital self-reports. We conducted an ANOVA comparing Fitbit compliance across studies and a series of linear regressions evaluating the relationships between baseline measures of symptom severity and compliance with Fitbit and digital self-report protocols. Fitbit compliance was defined as the proportion of weeks enrolled in which a participant wore the Fitbit 75% of the time.
Results: Both Fitbit and self-report showed a pattern of diminished compliance over time. BD participants showed lower Fitbit compliance than migraine participants [F(1, 67)=7.80, p=.007]. This remained true in sensitivity analyses correcting for sex and weeks in study. There was a significant negative relationship between baseline severity of manic symptoms and completion of bi-weekly self-reports in the BD sample and between baseline perceived stress completion of daily self-reports in the migraine sample.
Conclusions: Results support several conclusions. First, despite common perception that collecting passive sensor data is effortless for patients, lower compliance indicates that it may be more effortful than active data. Second, given that compliance appears to diminish over time, clinical applications of digital phenotyping should consider targeted observation windows when possible. Third, digital phenotyping may be more feasible and valid certain clinical populations and missing data may not be at random.