Beyond BMI: Machine Learning approach to identify superior obesity indicators for diabetes risk among Indian women

Abstract

Objective: To compare the effectiveness of five obesity indicators—body mass index (BMI), waist circumference (WC), waist-to-height ratio (WHtR), waist-to-hip ratio (WHpR), and waist-corrected BMI (wBMI, given by wBMI = WC * BMI) for predicting type-2 diabetes (T2D) among women of reproductive age in India using machine learning (ML) methods. Methods: This cross-sectional study is based on the National Family Health Survey dataset of 2021. Data corresponding to 613,752 non-pregnant women were analyzed. Diabetes status was determined using World Health Organization criteria based on self-reports and random blood glucose levels. Six datasets were constructed, five using one obesity indicator each, and the sixth one using all five obesity indicators. The performances of tree-based classifiers (C5.0, Random Forest, XGBoost) and penalized regression models (Ridge, LASSO, ElasticNet) were compared. The models were evaluated based on metrics such as sensitivity, specificity, and Area under the Precision-Recall Curve (AUPRC), which is the recommended metric for imbalanced data. Results: Diabetes prevalence was 3.83%. The wBMI, WC, and WHtR were good predictors. The wBMI (median AUPRC 0.105, median sensitivity 63.6%) performed slightly better than the WC (median AUPRC 0.102, median sensitivity 62.33%) and the WHtR (median AUPRC 0.102, median sensitivity 62.7%). Among the ML models, the LASSO and Ridge classifiers outperformed tree-based techniques. Conclusion: The results highlight the importance of fat distribution in the body in diabetes diagnosis. The wBMI, WC, and the WHtR could be effective alternatives to the BMI to be used in diabetes screening programs for Indian women. The superior performance of penalized regression and the use of non-invasive predictors offer a scalable approach for diabetes screening in low-resource settings, supporting targeted public health interventions. © The Author(s), under exclusive licence to Research Society for Study of Diabetes in India 2025.