Predicting diabetes complications from electronic health records visits

It is temporarily accepted

Diabetes has a major impact on millions of people around the world, leading to substantial morbidity, disability and mortality. Predicting diabetes-related complications from health records is important for early prevention and the development of effective treatment plans. This study introduces a new feature engineering approach to predict four different complications: diabetes, IE, retinopathy, chronic kidney disease, ischemic heart disease, and amputation. During classification model development, we utilize XGBoost feature selection methods and various monitored machine learning algorithms such as Random Forest, XGBoost, LogitBoost, Adaboost, and Decision Tree. These models were trained with synthetic electronic health records (EHRs) generated by an automated double by-product encoder. These EHRs represent nearly 1 million synthetic patients derived from a genuine cohort of 979,308 diabetes. The variables considered in the model were the age range with chronic diseases that occur during patient visits beginning with the onset of diabetes. Throughout the experiment, Xgboost and Random Forest achieved the best overall predictive performance. The final model tailored to each complication and trained using a functional engineering approach achieved accuracy between 69% and 77% and AUC between 77% and 84% using cross-validation. However, the partitioned verification approach has provided accuracy in between. 59% and 78%, while AUC is 66% to 85%. These findings imply that the performance of our method outweighs the performance of traditional bag sack approaches, highlighting the effectiveness of the approach in increasing model accuracy and robustness.