Machine learning for diabetes diagnosis: insights from the Erbil Diabetes Dataset and algorithmic performance

Machine learning technologies have brought significant operational improvements for disease diagnosis-related healthcare activities. Among various conditions, diabetes is particularly suited for prediction through historical and personalized data, which serve as a cornerstone of many machine learning applications. In this study, we present, for the first time, a newly developed, domain-specific diabetes research dataset, called Erbil Diabetes Dataset. The data were collected under the supervision of a medical professional at a laboratory in Erbil, Kurdistan Region of Iraq. The dataset contains twelve key characteristics which were captured from 662 people who visited the laboratory for check-ups. A standard procedure was employed to preprocess the features before presenting them to the diabetes research community for use. The performance evaluation of these algorithms on the specified dataset utilized five algorithms that included Random Forest (RF), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Xtreme Gradient Boost (XGBoost), and Decision Tree (DT). The evaluation process analyzed the performance results of algorithms through accuracy-recall measurements in addition to precision and F1-score metrics. The result shows KNN and XGBoost reaching outstanding performance values and predictive accuracy measures at 99.25% and 98.80% respectively. The accuracy levels of SVM decrease to 73.68% caused by their sensitivity to hyperparameter optimization. A statistical analysis using a one-way ANOVA test on F1-scores revealed a significant outcome (F = 558.51, p < 0.001), confirming that the differences in model performance were meaningful rather than due to random variation.