Trajectory-driven risk stratification of perinatal depressive symptoms based on an interpretable random forest model: a prospective cohort study.

BACKGROUND: Perinatal depression, as a global health challenge, is heterogeneous in its evolution, highlighting the need for improved trajectory-based prediction to enable timely and targeted intervention. METHODS: A longitudinal study was performed at four hospitals in China among 2303 perinatal women. Perinatal depressive symptoms were serially assessed from the second trimester to 42 days postpartum. Group-based trajectory modeling was used to identify heterogeneous trajectories, a random forest to develop a trajectory prediction model, and the SHAP method to quantify the contribution of each feature. RESULTS: 711 (30.9%) women experienced depressive symptoms at least once during the entire perinatal period. Five trajectories were identified: "well-adjusted" (14.3%), "chronic mild" (54.0%), "relieving" (25.4%), "moderate-to-severe depression persisting" (1.9%), and "suddenly deteriorating" (4.4%). A model developed using all factors to predict heterogeneous trajectories showed a modest improvement over that developed using relatively stable factors, with AUC values of 0.72 (95% CI: 0.68-0.76) and 0.68 (95% CI: 0.64-0.73). Key predictors were highly consistent across models and categories, with vulnerable-personality traits (mean |SHAP value|: 0.014-0.102) contributing most to the model output, followed by rumination (0.012-0.084), and psychological resilience (0.009-0.056). CONCLUSIONS: Nearly one-third of women experienced depressive symptoms at least once during the perinatal period, with heterogeneous patterns characterized by symptom of relief, persistence, or sudden deterioration. Perinatal women with concurrent features such as vulnerable personality and rumination are at high risk of unfavorable trajectories (i.e., persisting and suddenly deteriorating), calling for a concerted response from individuals, families, healthcare, and the public.