A recent study accepted for publication in Scientific Reports describes the development of a machine learning model designed to detect diabetes and assign cases to one of four diagnostic groups. The research team announced on Jul. 29 that their two-stage artificial intelligence framework demonstrated strong internal performance, but emphasized that its reliance on separate public datasets means further testing is needed before clinical use.
The model was trained using publicly available datasets related to diabetes, including the Pima Indians Diabetes Database maintained by the National Institute of Diabetes and Digestive and Kidney Diseases, as well as a dataset from the Kaggle repository. The approach used binary classification to identify diabetic versus non-diabetic records, followed by multiclass classification into prediabetes, type 1 diabetes (T1D), type 2 diabetes (T2D), or pancreatogenic/type 3c diabetes (T3cD). Inputs included age, body mass index, waist circumference, cholesterol levels, blood glucose levels, insulin levels, and a derived pancreatic-health index.
The researchers compared several machine learning algorithms such as logistic regression, decision trees, random forests, K-Nearest Neighbors (KNN), naive Bayes, and XGBoost. They selected XGBoost as their preferred classifier due to its ability to capture complex associations between features. Reported accuracy values for XGBoost ranged from 95.67% to 97%, though random forest achieved slightly higher accuracy in some tests. The team used Synthetic Minority Oversampling Technique on training data only and performed hyperparameter tuning for optimal results.
Analysis showed that blood glucose level was the most influential variable in predicting class labels according to both general feature importance measures and Local Interpretable Model-agnostic Explanations analysis. Age and cholesterol contributed secondarily in some classes, while other variables had less influence. The authors cautioned that these findings reflect patterns consistent with current clinical understanding but have not been independently validated in practice.
The study did not assess whether using this AI model would improve patient care or outcomes; nor did it evaluate quality of life or global burden reductions associated with its implementation. Researchers said additional studies should use clinically characterized datasets containing both status and subtype labels verified by biomarkers rather than derived variables.
