Abstract
Infant mortality is a major public health issue that is rooted in the larger problem of socio-economic and healthcare disparities. Deep learning techniques were employed in this study to predict infant mortality using data gathered via 2019 Ethiopia Demographic and Health Survey (EDHS). An Artificial Neural Network (ANN) was applied in four layers (input, two hidden, and output), where ReLU and sigmoid activation functions were applied, and the optimization was an Adam algorithm. This was tested for 96.2% accuracy, while a precision of the "immortal" class was 99%, with an AUC of 0.99, i.e., the model could differentiate between mortal and immortal cases with high confidence. Some of the key predictors are infant age, birth order, mother's age, and socio-economic predictors such as household size and wealth index. These findings affirm the capability of the deep learning paradigm in capturing complex and non-linear relationships that occur within health data to enable useful implications to policymakers and health workers alike. One limitation arises from the interpretability problem for algorithms and local generalizability; however, the study throws the spotlight on the potential of AI to reverse the trend towards reducing infant mortality in resource-scarce settings. Their inclusion in real-time surveillance systems with maximum emphasis on model interpretability for long-term uses is something future studies need to prioritize.