Wearable-based statistical modelling for predicting depressive symptom worsening

Background: Wearable actigraphy offers continuous monitoring of behavioural activity that may help predict depressive symptom trajectories. However, wearable prediction studies often rely on analyst defined temporal windows or high dimensional feature mining, increasing overfitting risk and limiting reproducibility. Standard rest activity metrics may provide a more interpretable modelling framework, particularly in small clinical samples where demographic imbalance can influence apparent predictive performance. Aims: We examined whether established nonparametric rest activity metrics could improve statistical modelling of short term depressive symptom worsening after adjustment for age and sex. Methods: We retrospectively analysed smartwatch accelerometer and gyroscope magnitude data from a clinical cohort with repeated depression assessments. Depressive worsening was defined as an increase in Hamilton Depression Rating Scale total score over follow-up. Sensor data were aligned to Korean standard time and summarised into five minute epochs. Rather than selecting investigator defined clock time windows, prediction models used standard nonparametric rest activity metrics from sleep wake research [1]: mean activity during the most active 10 hours (M10), mean activity during the least active 5 hours (L5), intradaily variability (IV), interdaily stability (IS), and relative amplitude (RA). Day to day variability measures were calculated for M10 and L5 activity. Hierarchical logistic regression estimated whether each metric added information beyond demographics. A baseline model included age and sex, followed by adjusted models adding each rest activity metric individually. Performance was assessed using repeated five fold cross validated area under the receiver operating characteristic curve (AUC). Associations were summarised using odds ratios per one standard deviation increment and likelihood ratio tests. Results: Of 177 participants with wearable recordings, 77 met standard metric coverage criteria; 12 showed depressive worsening. The worsening group was younger than the non-worsening group, and the age and sex baseline model showed substantial discrimination (AUC 0.769). Adding gyroscope derived IV improved adjusted modelling modestly and remained associated with worsening (adjusted AUC 0.778; odds ratio 2.57 per one standard deviation, p=0.041; likelihood ratio test p=0.018), indicating greater fragmentation of activity and rest transitions. Lower day to day variability of gyroscope M10 activity was also associated with worsening (adjusted AUC 0.773; odds ratio 0.22 per one standard deviation, p=0.044; likelihood ratio test p=0.013), consistent with reduced behavioural flexibility during the most active period. M10 and L5 timing metrics showed descriptive phase delay in unadjusted analyses, while fixed clock time nocturnal activity features showed stronger exploratory discrimination. These effects were less consistent after demographic adjustment, supporting their interpretation as complementary exploratory signals. Conclusions: Wearable-based statistical modelling with standardised actigraphy metrics identified clinically interpretable predictors of depressive worsening beyond demographic effects. Increased IV and reduced variability of M10 activity suggest that worsening may be characterised by fragmented daily activity patterns and reduced daytime behavioural flexibility. These findings support hypothesis driven rest activity metrics as candidate digital predictors for depressive symptom monitoring, while highlighting the need for prospective validation in larger cohorts.

References

1. Witting, W., Kwa, I.H., Eikelenboom, P., Mirmiran, M., Swaab, D.F., 1990. Alterations in the circadian rest-activity rhythm in aging and Alzheimer’s disease. Biol. Psychiatry 27 (6), 563-572.