MACHINE LEARNING PREDICTIVE MODEL FOR MATERNAL MORTALITY IN KENYA

Abstract

The rate of maternal mortality in Kenya is alarming and brings a primary public health concern, as it stands at 342 deaths per 100,000 live births, which exceeds the global mortality rate of 211. Approximately 5000 women lose their lives annually to pregnancy complications, and about 40 percent of these women have associated risk factors that include maternal age, parity, pre-existing conditions, home residence, and socio-economic factors. The remaining 60 percent is due to avoidable causes, which can be addressed through early interventions. This study sought to fix this prevailing problem by developing a predictive model of maternal mortality based on electronic medical records (EMRs) using a machine learning technique. A sample of 500,000 representative birth records from the Kenya Health Information System (KHIS) that was enclosed in more than 1.2 million births that were reported in the year 2022 was cleaned and analyzed using Python libraries, namely Pandas, NumPy, Seaborn, and Scikit-learn, to perform data processing, computation, visualization, and model training, respectively. Based on the principles of Eric Topol of predictive care in machine learning, which implies the improvement of the quality of diagnostics, the possibility to intervene before the beginning of disease progression, and the ability to make clinical decisions based on data, four models were implemented and tested, namely Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost).The accuracy, precision, recall, F1-score, and ROC-AUC were applied to measure the performance of models. XGBoost performed better than all other models, with the result of 0.89 accuracy, 0.87 precision score, 0.92 recall, 0.89 F1-score, and 0.90 ROC-AUC value, thus indicating the efficacy of an ensemble learning categorization in learning complex non-linearities on maternal health data. In the study, the authors describe the promise of predictive modeling to make early-stage risk detection, facilitate timely interventions, improve allocations of resources, and support evidence-based policymaking. The results back the idea of implementing machine learning in Kenya's hospital-based decision support systems (DSS) to improve maternal health.