An Assessment of Neural Network Efficacy in Hospital Readmission Prediction for Diabetic Patients

Predicting hospital readmission among diabetic patients is an important aspect of improving the conditions of the patients and ensuring healthcare resources are optimally used. This paper seeks to find how effective neural networks in predicting hospital readmissions are by increasing the accuracy. The UCI Machine Learning Repository dataset was pre-processed for missing values, outliers, and categorical encoding. Class imbalance was also handled using the SMOTE technique. The final developed neural network with Adam optimizer and binary cross-entropy loss function found an accuracy of 93% from a balanced test set and 91% for the original imbalanced dataset. Class 0 (no readmission) found precision of 0.91; class 1 (readmission) found precision of 0.70. Recall for class 0 was 1.00, and for class 1, it was 0.01. The AUC-ROC for the imbalanced dataset was 0.65. Such findings highlight the potential of the model in recognizing patients at high risk and underline its utility, at the same time as it shows the model's boundaries. The developed research fosters knowledge on how neural networks can find application in this field and will support healthcare workers in developing preventive interventions, while also revealing areas for possible further refinement.