A Novel Approach for Predicting Diabetic Readmissions Using Feature Selection and Optimization Techniques

Abstract

Forecasting readmissions in diabetic patients is critical for developing patient services and healthcare systems. This work proposes a new classification model simplifying model building to predict patients' readmission risk stratified into short-term (up to 30 days) and long-term (beyond 30 days) cohorts. This model uses Grey Relational Analysis (GRA) for feature selection and utilizes Grey Wolf Optimization (GWO) to enhance model parameters for better accuracy. The UCI diabetic readmission dataset used contains clinical attributes including age, BMI, glucose and insulin levels, HbA1c, total cholesterol, triglycerides, and newer markers such as ‘new’ cholesterol and ‘new’ triglycerides. The study first applies GRA computing Grey Relational Coefficients GRC for every described feature concerning the readmission status. It then ranks features in terms of their contribution for the classification task to build a model. Then GWO optimizes the feature coefficients together with the discriminant coefficient which increases the model's predictive accuracy. The method proposed, which incorporates GRA for feature selection and GWO for model tuning outperformed the other classifiers that used GRA, achieving the highest accuracy of 98.6% when using a decision tree, and 97.3%, 97.6% with gradient boosting. The results show that GRA is effective in feature selection while GWO optimizes the classification model. The approach developed in the study for estimating the risks of diabetic readmissions is quite practical. Adding lifestyle factors, genetic indicators, and real-time data could enhance the model's adaptability to different scenarios, improving its utility in healthcare systems.