An Uncertainty-Aware Enhanced MLP Framework with Hybrid Feature Selection for Early Prediction of Postpartum Depression Risk Using Clinical Data

Abstract

Postpartum Depression (PPD) is a serious maternal mental health problem that has traditionally been undiagnosed by a lack of initial screening systems and a dependence on clinical judgment of the subject. The main issue, which is discussed in the given study, is the absence of reliable, interpretative, and datagrounded anticipatory frameworks of early risk detection of PPD. The aim of the study is to create a better deep learning(DL) system capable of ensuring accurate prediction of PPD based on structured clinical information. A sample of 3,000 postpartum records including demographic, psychological and obstetric characteristics is employed. The methodology suggested consists of ample data argumentation, derivation of EPDS scores, the hybrid selection of features, and a fine-tuned Multi-Layer Perceptron (MLP) model with embedding layers, residual connections, GELU activation, AdamW optimization, and Monte Carlo dropout to estimate uncertainty. The experimental findings show that the proposed model is better performing with an AUROC of 0.929, precision of 0.887, an accuracy of 0.880 and a recall of 0.885, achieving better results than the traditional machine learning(ML) models. The paper concludes that the suggested framework can greatly improve the predictive accuracy and present a clinically interpretable instrument on early PPD risk identification, allowing the timely intervention and better health outcomes in mothers.