Leakage-Controlled and Survey-Weighted Machine Learning for Neonatal Mortality Risk Prediction Using NFHS-5 Data

Area of Science:

• Public Health
• Machine Learning
• Biostatistics

Background:

• Neonatal mortality rates vary significantly across Indian states.
• Existing prediction models struggle with imbalanced data, data leakage, and complex survey designs.
• This research addresses these limitations using advanced machine learning techniques.

Purpose of the Study:

• To develop and evaluate survey-aware machine learning models for neonatal mortality risk prediction.
• To compare the performance of different machine learning algorithms using National Family Health Survey (NFHS-5) data.
• To assess model calibration, generalizability, and state-level transportability.

Main Methods:

• Analysis of NFHS-5 data from 33,338 children in Bihar, Chhattisgarh, and Uttarakhand.
• Application of survey-aware techniques: grouped cross-validation, sampling weights, fold-contained preprocessing, and feature augmentation.
• Comparison of logistic regression, random forest, histogram gradient boosting (HGB), and artificial neural networks, with PR-AUC as the primary metric.

Main Results:

• Histogram Gradient Boosting (HGB) demonstrated the highest predictive performance (ROC-AUC 0.755, PR-AUC 0.202 on the test set).
• While HGB showed high sensitivity (0.813) and NPV (0.989), PPV was low (0.063).
• Certain methods like clustering and augmentation provided minimal additional value; state-level performance varied.

Conclusions:

• Survey-weighted HGB offers the most robust predictive performance for neonatal mortality risk.
• The model's low positive predictive value and variable state-level results limit its application to preliminary, low-cost screening.
• Further prospective validation is essential before considering deployment in clinical or public health settings.