Dynamic prediction and identification of cases at risk of relapse following completion of low-intensity cognitive behavioural therapy.

Objective: Low-intensity cognitive behavioural therapy (LiCBT) can help to alleviate acute symptoms of depression and anxiety, but some patients relapse after completing treatment. Little is known regarding relapse risk factors, limiting our ability to predict its occurrence. Therefore, this study aimed to develop a dynamic prediction tool to identify cases at high risk of relapse. Method: Data from a longitudinal cohort study of LiCBT patients was analysed using a machine learning approach (XGBoost). The sample included n = 317 treatment completers who were followed-up monthly for 12 months (n = 223 relapsed; 70%). An ensemble of XGBoost algorithms was developed in order to predict and adjust the estimated risk of relapse (vs maintained remission) in a dynamic way, at four separate time-points over the course of a patient's journey. Results: Indices of predictive accuracy in a cross-validation design indicated adequate generalizability (AUC range = 0.72-0.84; PPV range = 71.2-75.3%; NPV range = 56.0-74.8%). Younger age, unemployment, (non-)linear treatment responses, and residual symptoms were identified as important predictors. Discussion: It is possible to identify cases at risk of relapse and predictive accuracy improves over time as new information is collected. Early identification coupled with targeted relapse prevention could considerably improve the longer-term effectiveness of LiCBT.