Clinical feature grouping-based FT-Transformer for postpartum depression risk prediction: a longitudinal cohort study...

Postpartum depression (PPD) poses risks to maternal health and family relationships, making identification of high-risk individuals important. Most PPD prediction models treat clinical variables as an undifferentiated set and rarely use the clinical grouping structure of risk factors. Using perinatal and postpartum data from 1138 women, including clinical characteristics, biochemical indicators, and psychological assessments such as the Eysenck Personality Questionnaire (EPQ) and Beck Depression Inventory (BDI), with PPD defined as an Edinburgh Postnatal Depression Scale (EPDS) score of at least 11 at 6 weeks postpartum, we incorporated clinical feature grouping into Feature Tokenizer + Transformer (FT-Transformer) and developed Clinical-Group FT for PPD risk prediction. Clinical-Group FT divides predictors into four clinical feature groups, learns within-group representations, and integrates information through inter-group cross-attention and gated fusion. Models were evaluated using an 8:2 stratified split and five-fold cross-validation. F1 score was selected as the primary metric because PPD screening requires identifying positive cases while limiting false positives. Paired bootstrap testing was used for F1 comparisons. On the independent test set, Clinical-Group FT achieved an accuracy (ACC) of 0.851, an area under the curve (AUC) of 0.887, and an F1 score of 0.734, demonstrating the best performance among the nine evaluated models. Ablation experiments supported the contribution of the proposed components. Dual-level interpretability analyses highlighted mother-in-law's care in postpartum, antenatal depression, and the neuroticism dimension of the EPQ as important contributors. These findings indicate that Clinical-Group FT improves PPD risk prediction while providing interpretable information for screening support.