JMIR Publications

ABSTRACT

Background:

Digital phenotyping — the use of personal digital devices to capture real-world behavioral and physiological data — holds promise for measuring patient function over time. However, missing data remains a pervasive challenge in longitudinal studies where missingness may obscure clinically relevant signals. Existing statistical methods focus on group-level inference and offer limited guidance on how causes of missingness should inform clinical interpretation and decision-making.

Objective:

This study aimed to (1) characterize data missingness across multiple timescales in a longitudinal digital phenotyping study of patients with chronic pain, (2) evaluate the effect of imputation on clinical inference, and (3) propose a practical framework for categorizing and responding to missing data in clinical digital phenotyping research.

Methods:

We analyzed data from 85 patients with chronic musculoskeletal pain (mean age 55.2 years, SD 15.7; 51 female, 32 male, 1 transgender) recruited from the Pain Intervention and Digital Research Program. Active data (PROMIS-29 surveys, daily pain scores) and passive data (accelerometer, GPS) were collected via the Beiwe Research Platform over 180 days. Data completeness was computed at day, hour, and minute levels. Linear mixed-effects models assessed associations between daily missingness and Forest-derived summary measures (cadence, home time, significant locations). Cumulative Link Mixed Models and linear mixed models with autoregressive error structures evaluated associations between PROMIS domain scores and digital measures, adjusting for age, sex, race, and within-subject correlation. Complete-case analyses were compared against multiple imputation (predictive mean matching, proportional odds logistic regression, and MidasTouch) using 100 imputed datasets combined via Rubin's rules.

Results:

Median accelerometer completeness was 60% at the day level, 37% at the hour level, and 26% at the minute level; GPS completeness followed a similar pattern (57%, 34%, and 5%, respectively). Cadence showed no significant association with missingness (false discovery rate [FDR]-adjusted P=.32). The number of significant locations declined modestly with higher missingness (beta=-0.073 per 10 percentage points; FDR-adjusted P<.001). In complete-case analysis, higher cadence was associated with lower depression scores (95% CI -2.09 to -0.28; P=.01); this association was attenuated after multiple imputation (P=.13). Older participants remained enrolled longer (hazard ratio 0.979 per year; P=.02) but were less engaged while enrolled (odds ratio 0.962; P=.004). Race and sex were not significantly associated with engagement or retention.

Conclusions:

Data missingness in digital phenotyping varies substantially by the timescale at which it is assessed, and imputation choices can alter clinical interpretations. We propose the Triage and Response for Interpreting Missingness (TRIM) framework, which categorizes causes of missing data into Technology Failure, Clinically Relevant events, and Extraneous life events, each requiring distinct operational and analytical responses depending on the clinical context. TRIM provides a shared vocabulary for clinicians, statisticians, and engineers to ensure that missing data is meaningfully interpreted rather than merely imputed.