Comparative Evaluation of Tree-Based Algorithms for Postpartum Mental Health Risk: a Methodological Benchmark

Abstract

Gradient-boosting classifiers now underpin most screening tools proposed for postpartum mental health, but published comparisons rarely separate what the modelling pipeline contributes from what the clinical construct itself contributes. Here we benchmark XGBoost, CatBoost, and LightGBM inside a single leakage-controlled pipeline, asking whether performance gaps trace back to identifiable structural properties rather than inconsistent data handling. A public postpartum survey dataset of 1,503 records was split 80:20 into development and test sets, with encoding and imputation fitted on the development set alone, and each classifier tuned via stratified five-fold cross-validation over a shared search space, isolating architecture as the sole variable. LightGBM's histogram-based, leaf-wise growth produced the strongest held-out discrimination (accuracy 88.4%, weighted F1 88.4%, balanced accuracy 87.8%, macro F1 88.0%), ahead of XGBoost (86.0%) and CatBoost (82.4%), with its advantage concentrated in the Moderate and High categories (81 of 92 High-category records correctly classified, versus 80 and 76). The selected classifier was integrated into a four-layer architecture, preprocessing, predictive inference, dialogue management, and deterministic recommendation, keeping the model's output separate from user-facing text and independently auditable. We report the full search space and the structural mechanisms plausibly explaining the ranking, and state plainly that training time, memory, and inference latency were not measured, naming this as required future work under TRIPOD+AI. The results demonstrate a reproducible method for comparing boosting architectures on a fixed screening task, they do not establish diagnostic validity, calibrated suicide-risk estimation, or clinical readiness, which require external validation beyond this internal benchmark.