Investigating the impact of missing value imputation methods on the prediction of diabetes using machine learning

Diabetes poses significant challenges due to its prevalence and the potential consequences of inaccurate or delayed diagnosis. This study focuses on enhancing prediction reliability to mitigate such risks. Initially, it identifies diabetes-related factors through correlation analysis with the target variable and implements models to address missing data. Subsequently, various imputation methods including CART, GMM, and RFR are employed to evaluate these factors. Results from each imputation scenario inform the selection of the most effective method. The study then employs ensemble algorithms like AdaBoost, Bagging, Gradient Boosting, and RF to enhance classification model accuracy. Further refinement is achieved by optimizing hyper-parameters through grid search. Evaluation involves comparing model predictions with those of medical professionals to assess accuracy. The findings reveal superior performance of optimized machine learning models over human predictions, indicating potential for improved diagnosis accuracy and reduced medical errors. This research contributes to advancing predictive modeling in diabetes diagnosis, offering prospects for enhanced community health and reduced socioeconomic burdens.