Machine Learning Approaches to Identify Neonates at Risk for Post-Discharge Mortality in Dar es Sala

Music City Center 

Machine learning (ML) can increase discriminatory value in risk assessment tools compared to traditional regression. We explored the performance of ML models, compared to a previously derived logistic regression model (area under the curve [AUC]=0.77, 10 variables), for predicting all-cause mortality within 60 days post-discharge among neonates from two national referral hospitals in sub-Saharan Africa.
In a prospective cohort of 2,294 neonates (3% mortality rate), data were randomly split (80% training, 20% testing). We addressed class imbalance with Synthetic Minority Oversampling and selected variables via minimum-Redundancy maximum-Relevance. We trained random forest, XGBoost, hist gradient boosting, support vector machine (SVM), and neural network models, optimizing hyperparameters via 5-fold cross-validation.
Hist gradient, random forest, and XGBoost achieved AUCs of 0.99 with six variables. Neural network (AUC=0.97) required eight, and SVM (AUC=0.89) required 17 but was computationally heavy. ML models outperformed logistic regression (p<0.001). Selecting parsimonious, high-accuracy, low-cost models are key for feasible clinical implementation.

Machine learning

Prediction modeling

Logistic regression

Model performance

Risk prediction 

Section on Statistics in Epidemiology