Related Experiment Videos
Evaluating machine learning models for predicting postpartum depression symptoms using mississippi pregnancy risk
Hailay Berihu Abebe1, Negasi Tsegay2, Girmay Berhie1
1School of AI, Informatics and Analytics, Jackson State University, Jackson, MS, the United States of America; Mississippi Delta Center of Excellence in Maternal Health, USA.
Background:
Postpartum depression (PPD) is a significant public health concern with profound effects on maternal well-being and infant development. Building a highly accurate predictive model for PPD symptoms using data-driven approaches can facilitate earlier screening and improve intervention outcomes.
Objective:
The objectives of this study are to evaluate the performance of machine learning (ML) models for predicting self-reported PPD symptoms, identify the most important features contributing to these predictions, and assess model performance when using a reduced set of features, all based on Mississippi PRAMS 2016-2022 data.
Methods:
After preprocessing, including handling missing values, recoding variables, feature scaling, and addressing data imbalance, the Mississippi PRAMS 2016-2022 dataset from the Mississippi State Department of Health (MSDH) was prepared for analysis. Ten ML models were initially trained using 138 input features. The models were subsequently retrained using a reduced set of important features to evaluate predictive performance and compare the results with previous studies.
Results:
Binary classification of PPD symptoms (screening-positive versus healthy) was performed using ten ML algorithms based on 138 multivariate features. Using the 40 most important features identified by the best-performing algorithm (gradient boosting machine (GBM)), the ten models were re-evaluated. The highest area under the curve (AUC) values were achieved by GBM (0.807), histogram-based gradient boosting machine (HistGBM) (0.797), AdaBoost (0.793), and Random Forest (RF) (0.792). The GBM model achieved a sensitivity of 0.491 and a specificity of 0.960. These results compare favorably with those reported in a previous PRAMS-based study while using nearly four times fewer input features.
Conclusions:
Training ML models with a reduced set of top-ranked features improves efficiency and interpretability while minimizing the risk of overfitting. Feature selection based on the best-performing model maintained predictive performance and model efficiency, particularly for ensemble-based algorithms. These findings support the use of ML models for predicting screening-positive PPD symptoms to assist population-level screening and maternal mental health interventions.