Machine Learning Prediction of Preterm Birth: An Analysis of Facility-Based Paper Health Records in Uganda

Abstract Purpose: Preterm Birth (PTB) is one of the leading causes of neonatal mortality in Uganda. Machine Learning (ML) can be used to identify women at risk of PTB in time for medical intervention and adequate preparation by mothers. Methods: We utilized data from paper-based maternal health records at Kawempe National Referral Hospital, Uganda. A case-control method was employed, where for every woman who experienced a PTB, a woman without PTB and delivered in the same day was selected as a control. Treatment of missing data was done using Random Forest imputation. Variable Importance was analyzed using Random Forest. The following classification methods were applied in the prediction of PTB: Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Naïve Bayes (NB). Performance of methods was investigated using prediction accuracy, sensitivity, and specificity. Results: 1,540 women were included in the study, where 770 women had experienced PTB, and 770 women formed the controls. According to variable importance analysis, number of antenatal care visits had the biggest impact on PTB. SVM had the highest accuracy in predicting PTB at 64% (sensitivity 64% and sensitivity 63%). Conclusions: Prediction of PTB using paper-based records in a developing country yielded similar results to studies done using electronic health records in developed countries. The predictive power could be low in this study due to fewer variables available from routinely collected ANC data. The inclusion of significant variables in the maternal records could potentially increase predictive power.