Opportunistic application of machine learning for bone health stratification in type 2 diabetes mellitus

Purpose The primary objective of this study was to compare the performance of various machine learning (ML) models in the opportunistic screening of abnormal bone mineral density (BMD) among type 2 diabetes mellitus (T2DM) patients using routine non-contrast-enhanced chest CT. Materials and methods A total of 249 T2DM patients (age range, 40–97 years; mean ± standard deviation, 60.93 ± 11.49 years) who underwent chest CT with quantitative CT (QCT) analysis were retrospectively enrolled between January 2021 and June 2024. Patients were randomly split at a 7:3 ratio into a training set (n=175) and an internal test set (n=74). Semi-automatic segmentation of the L1 vertebral body was performed using 3D Slicer, and 2273 radiomic features were extracted via the Research Portal V1.1 workstation. Six radiomics-based ML models were constructed; their predictive performance and stability for distinguishing normal versus abnormal BMD were evaluated and compared. Results Among the T2DM patients included, the incidence of abnormal BMD was 40.56% (101/249). Fifteen radiomic features were selected using the minimum redundancy-maximum relevance (MRMR) algorithm and incorporated into six ML models (RF, SVMRadial, SVMLinear, NNet, xgbLinear and KNNs). In nested cross-validation, the SVMLinear model achieved the better predictive performance (accuracy: 0.906 ± 0.032; area under the curve [AUC]: 0.945 ± 0.028), followed by the NNet model (accuracy: 0.906 ± 0.031; AUC: 0.941 ± 0.029). In terms of stability, the SVMLinear model exhibited greater robustness against data perturbation (relative standard deviation [RSD] of mean accuracy: 3.54%; RSD of mean AUC: 3.04%). Conclusions The ML models enable opportunistic prediction of BMD status from chest CT in patients with T2DM. Specifically, the L1-vertebra radiomics-based SVMLinear model achieves a favorable balance of discrimination ability and stability, and may serve as a useful tool for opportunistic clinical screening for abnormal BMD in this patient population.