<p>Prolonged grief disorder (PGD), characterised by persistent and disabling grief, is officially recognised as a mental disorder. Accurate prediction of prolonged grief (PG) symptomatology is vital to intervention efforts. We trained machine learning models to predict PG severity using cross-sectional data from the MARBLES archive (<i>N</i> = 5570). With explainable AI (XAI), we examined linear and nonlinear relationships, individual differences, and interactions between predictors, including seven cognitive-behavioural variables (i.e., a sense of unrealness, negative cognitions about the self, life, and the future, threatening interpretations of grief, and anxious and depressive avoidance) alongside socio-demographic characteristics, loss characteristics, and concurrent psychological symptoms (i.e., depressive and posttraumatic stress symptoms). Among the evaluated models, XGBoost performed best. Cognitive-behavioural variables, particularly a sense of unrealness, were the most important predictors. Analyses revealed both linear (e.g., a sense of unrealness) and nonlinear (e.g., age) associations of predictors with PG severity and identified key interactions between predictors, mostly involving depressive avoidance and threatening interpretations of grief. Findings suggest the value of cognitive-behavioural variables in predicting PG and XAI’s role in informing targeted intervention (e.g., behavioural activation).</p>