Abstract
BACKGROUND: Panic disorder (PD) is a chronic and impairing anxiety disorder. Individuals with more sleep disturbances might be predisposed to nine-year PD chronicity. However, linearity assumptions, small predictor sets, and analytic and design limitations have hindered optimal identification of which sleep disturbance variables are distal risk factors for PD chronicity. We thus used machine learning (ML) to predict nine-year PD chronicity using high-dimensional data. METHOD: Community-dwelling adults (N = 1054) completed clinical interviews, self-reports, and seven-day sleep actigraphy at Wave 1 (W1) and the same clinical interview at Wave 2 (W2) nine years later. The baseline data comprised 43 actigraphy, self-reported sleep disturbances, clinical, and demographic variables. Seven ML models were examined. Gradient boosting machine (GBM) was the best-performing algorithm. PD chronicity was defined as the presence of a PD diagnosis at both W1 and W2. RESULTS: The GBM accurately predicted PD chronicity (area under the receiver operating characteristic curve [AUC] =.764). Shapley additive explanation analysis showed that the top W1 predictors of PD chronicity were comorbid major depressive disorder, low healthcare utilization, sleep medication use, lengthier wake after sleep onset, and sleep-wake circadian disruptions based on actigraphy and self-reports. Lower household income and younger age were also top predictors. Additionally, the final multivariate model was well-calibrated. CONCLUSIONS: As proposed in our sleep-panic nexus theory, actigraphy and subjective sleep disturbances have essential prognostic value in predicting long-term PD chronicity. Harnessing ML facilitates accurate prediction by identifying complex, nonlinear relations across high-dimensional datasets, possibly improving prevention and treatment tailoring.