JMIR Publications

ABSTRACT

Background:

Insulin resistance (IR) is central to diabetes pathogenesis and a risk factor for chronic diseases.

Objective:

This study aims to develop and validate a machine learning (ML) prediction model for IR in a non-diabetic population, using low-cost diagnostic indicators and questionnaire surveys.

Methods:

This retrospective study analyzed data from individuals who underwent physical exams and completed surveys at the Health Management Center of Xiangya Third Hospital, Central South University, from January 2018 to August 2022. Insulin sensitivity was assessed using the HOMA-IR method. A total of 17,287 non-diabetic participants aged 18 and above were included for model training with 90 features. Five ML algorithms—Random Forest (RF), LightGBM, XGBoost, GBDT, and CatBoost—were used. Model optimization included resampling, feature selection, and hyperparameter tuning. Performance was evaluated using F1 score, accuracy, sensitivity, specificity, area under the curve (AUC), and Kappa value. SHAP analysis was applied to assess feature importance. A retrospective cohort followed for five years was used for time-to-event analysis with Kaplan-Meier survival curves to validate the model’s ability to predict diabetes incidence.

Results:

Data from 16,411 non-diabetic individuals were analyzed. The dataset was split 80% for training and 20% for testing. The final model included 34 lifestyle-related questionnaire features and 17 biochemical markers. The LightGBM model showed the best IR prediction performance with an accuracy of 0.7542, sensitivity of 0.6639, specificity of 0.7642, F1 score of 0.6748, Kappa value of 0.3741, and AUC of 0.8456. In the cohort, the high-risk IR group exhibited a significantly higher 5-year cumulative incidence of diabetes.

Conclusions:

The LightGBM model effectively predicts IR status in non-diabetic individuals and identifies those at high risk of progressing to diabetes, supporting its clinical utility.