AUTONOMOUS DIABETES DIAGNOSIS USING XGBOOST AND DECISION TREE CLASSIFIERS ON ROUTINE CLINICAL DATA

Abstract

Diabetes mellitus affects an estimated 537 million adults worldwide, however, many cases go undiagnosed until complications arise. This results in a confirmation that depends on the periodic testing of tissues in a laboratory, whose results have to be interpreted manually, leading to a delay in diagnosis in cases of limited availability of specialists. The implementation of a machine-learning autonomous diagnostic system that assesses the information about patients, which can be found in the records of internal organs and laboratory work, has been highlighted in the present article. The machine-learning diagnostic system is organized as a pipeline, which consists of the processes of data-eaned, data allocation, model fitting, and diagnosis output as a class label with a probability score; thereby, the system returns not only a label of a class but also the corresponding probability score as a measure of the system's confidence. The study used 27,551 anonymized patient records from the publicly available Kaggle Diabetes Prediction Dataset, described by age, gender, hypertension, heart disease, smoking history, body mass index, glycated hemoglobin (HbA1c) and blood glucose, with a diabetic prevalence of about 9%. After duplicate removal, categorical encoding and standardization, a Decision Tree and an XGBoost classifier were trained on 70% of the records and tested on the remaining 8,266. XGBoost achieved an accuracy of 96.7%, a precision of 93.5%, a recall of 69.0% and an F1-score of 79.4%, against 94.9%, 71.6%, 72.8% and 72.2% for the Decision Tree, and it reduced false-positive diagnoses from 217 to 36. In a complementary experiment, a random forest trained with SMOTE oversampling and random under sampling raised recall to 83.3%, and HbA1c and blood glucose together accounted for three-quarters of its feature importance. These findings show that tree-based models built on routine indicators can screen for diabetes accurately with few false alarms, and that rebalancing the minority class is the most effective lever for improving sensitivity before clinical deployment.