When Users Log Mood After Assessment: Temporal-Lag, Discriminant Validity, and the Simplicity Paradox of Single-Label Ecological Momentary Mood Tracking in a Saudi Arabic-Language Digital Mental Health Cohort-A Diagnostic-Accuracy Study.
文献摘要
Background : Digital mental health platforms increasingly deploy self-directed single-label mood tracking as an ecological momentary assessment (EMA) tool, often framed as a screening indicator for depression or anxiety. Whether such naturalistic mood logs accurately detect screening-positive PHQ-9 and GAD-7 scores has not been tested at scale against validated Arabic-language reference-standard instruments. Objectives: To evaluate, retrospectively, how accurately naturalistic single-label Arabic mood tracking detects screening-positive scores on the validated Arabic PHQ-9 and Arabic GAD-7 administered on the same platform, and to characterise the convergent-validity signal on continuous scores. Methods : This was a retrospective linked-records study of a Saudi Arabic-language commercial digital mental health platform. From 25,555 adult mood-logging users (aged 18 years or older), we assembled three intersection cohorts: depression (n = 4236), anxiety (n = 2449), and both-scales (n = 1825). Nine Arabic mood labels were categorised a priori as negative, positive, or neutral. Operating characteristics with bootstrap 95% confidence intervals were computed for four primary indices; thirteen comparator specifications and four machine-learning models were compared against the best-performing index. AUC differences were evaluated with a DeLong-equivalent bootstrap test. Reporting followed STARD. Results : The best-performing index (proportion of negative logs) yielded an AUC of 0.624 (95% CI 0.604-0.642) for screening-positive detection of depression and 0.652 (0.629-0.676) for anxiety, both below the conventional 0.80 benchmark for a clinically useful screening test. Youden-optimal sensitivity and specificity were 0.522 (0.506-0.538) and 0.712 (0.679-0.741) for depression and 0.569 (0.547-0.589) and 0.706 (0.670-0.743) for anxiety. Positive likelihood ratios of 1.81 and 1.93 fell below the 2-5 range that is conventionally associated with even small clinically meaningful shifts in post-test probability. Spearman correlations with continuous scores were modest but stable across bootstrap resamples (ρ = 0.232 for depression, 0.268 for anxiety; both p p = 0.005) but not for the anxiety-specific single label (Z = 0.32, p = 0.75), reflecting a 2.24% anxiety-label prevalence on this platform. Third, we observed a simplicity-paradox pattern in which added composite complexity either failed to raise cross-validated AUC or destroyed discriminant validity. Conclusions : On this platform, single-label Arabic mood tracking shows a modest but statistically reliable retrospective association with screening-positive status on the validated Arabic PHQ-9 and Arabic GAD-7 that does not reach the conventional 0.80 AUC benchmark for a clinically useful screening test; it should not substitute for these validated instruments. It is best treated as a longitudinal engagement and self-monitoring indicator rather than a stand-alone screening tool.