The use of Machine Learning to predict adverse birth outcomes: Empirical real world evidence from a human cohort study in Adama, Ethiopia

Sammanfattning: Since Ethiopia has a high number of recorded adverse birth outcomes, the city of Adama was subjected to a study (Flanagan et al., 2022) that gathered data from 2085 pregnancies. This thesis utilizes that data to investigate the usage of machine learning in environmental epidemiology. Using the classification methods logistic regression, random forest, support vector classifier, and k-nearest neighbors, two different sampling methods were implemented to handle the imbalanced dataset. The original imbalanced dataset performed worst though similar to the undersampled dataset. Oversampling the dataset with SMOTE yielded the best result with the random forest classifier and had an AUC score of 0.72 and an f1-score of 0.85. With further work, more data, and higher evaluation scores, machine learning may be a way to implement preventable medicine in Adama, Ethiopia. However, more research is needed, especially with larger study populations, to improve the accuracy of these models and find the most important features to analyze for this region.