JMIR Publications

ABSTRACT

Background:

With rising mental health complaints, the preventive monitoring of mental health becomes increasingly important. Different facets of mental health, such as perceived stress and mental wellbeing, have often been assessed with self-report questionnaires, administered at sporadic points in time. Ecological Momentary Assessment (EMA) may provide more continuous insights into these constructs. Research suggests that momentary affective states correlate with long-term mental health outcomes and that temporal EMA features of affective states (eg, average, variability, inertia) may be indicative of mental health complaints. However, findings have been mixed due to variability across settings and individuals. Long-term prediction performance has also not been thoroughly explored yet, partly due to the scarcity of longitudinal daily assessment data.

Objective:

This paper examined the predictive accuracy of temporal EMA features for changes in mental health and investigated how the number of EMA data points (time windows) influences prediction performance.

Methods:

Fifteen PhD candidates in the Netherlands, recruited via e-mail and other media, completed daily EMA of affect (valence, arousal; N=3953) and retrospective, validated monthly questionnaires (N=153) over 9 months. Retrospective outcomes included distress, perceived stress, personal accomplishment, exhaustion, and mental well-being. Temporal EMA features (average, variability, inertia, coefficient of variation) were aggregated to align with the monthly questionnaires (N=153) and used to predict changes in retrospective outcomes (R2pred) with leave-one-subject-out (LOSO) linear prediction models. In addition, within-subjects linear regressions were conducted to examine the explained in-sample variance. Prediction windows ranged from 3 to 30 days before questionnaire completion.

Results:

The linear LOSO models yielded the highest median R2pred with a 3-day window for Distress (R2pred=0.44), Mental wellbeing (R2pred=0.25), and Personal accomplishment (R2pred=0.20), versus a 14-day window for Perceived stress (R2pred=0.16) and a 16-day window for Exhaustion (R2pred=0.19). Results did not show monotonic trends across time in predictive accuracy nor a consistently optimal window length. Within-subjects linear regressions indicated similar findings, with variation in performance within and across prediction windows.

Conclusions:

Preliminary evidence suggests that temporal EMA features can predict changes in retrospective outcomes of mental health at a limited scale, with performance varying by window length. This discrepancy between momentary and retrospective assessments aligns with prior research on recall and may reflect individual differences in recall strategies. Findings suggest that shorter time windows between EMA and retrospective assessments or EMA alone, ideally combined with continuous objective measures, may offer a more detailed picture of changes in a person’s mental health.