Optimal Model Selection Methodology for Machine Learning-Based Diabetes Risk Prediction

Machine learning has been shown to have widespread application in the accurate and consistent determination of disease risk for multiple diverse disease types. Among such types of machine learning models, tree-based algorithms are popular for their simplicity and effectiveness. Meta-estimation models have also demonstrated effectiveness for multiple such scenarios. The present work consequently examines and compares the performances of multiple tree-based models with significant difference in model complexity, in accurately predicting diabetes in prospective patients. Evaluation metrics such as accuracy, precision, recall and F1-score are employed to determine the optimal model for a given dataset among a handful of such models. For determination of model optimality in the context of a specific dataset, the variation in performance measures is also observed with change in factors such as the ratio between training and testing dataset sizes. Based on such assessments, optimal model performance can be isolated and such optimal models identified for diverse use cases.