An Algorithm for Identifying Microvascular Complications in Type 1 Diabetes in the Clinical Setting: Development and...

Background: To facilitate appropriate medical stratification in clinical settings, this research leveraged externally validated machine-learning models to screen for microvascular complications specifically in the type 1 diabetes (T1D) population. Methods: By utilizing a cohort of 433 individuals with T1D from Peking University People's Hospital, we implemented eXtreme Gradient Boosting-based modeling combined with internal cross-validation to identify 3 microvascular complications: diabetic retinopathy (DR), diabetic nephropathy (DN), and diabetic peripheral neuropathy (DPN). The models were subsequently subjected to 2 external validations, utilizing datasets from patients with T1D hospitalized at Tangshan Gongren Hospital and Xingtai People's Hospital in Hebei Province (N = 126, named as external validation 1), as well as from the Chinese T1D Comprehensive Care Pathway program (N = 929, named as external validation 2). To evaluate the discriminative power of the developed models, we employed the area under the receiver operating characteristic curve (AUROC) as the primary evaluation metric. Results: The development set (N = 433), with a median age of 52.0 years and 52.7% female, exhibited prevalence rates of 24.2% for DR, 13.6% for DN, and 45.3% for DPN. During internal validation, the models achieved AUROCs of 0.773 (95% confidence interval [CI], 0.704 to 0.841) for DR, 0.803 (95% CI, 0.745 to 0.862) for DN, and 0.741 (95% CI, 0.730 to 0.753) for DPN. In external validation 1 (N = 126), the AUROCs were 0.850 (95% CI, 0.761 to 0.924) for DR, 0.727 (95% CI, 0.617 to 0.829) for DN, and 0.815 (95% CI, 0.732 to 0.886) for DPN. In external validation 2 (N = 663), the AUROCs were 0.822 (95% CI, 0.760 to 0.877) for DR, 0.832 (95% CI, 0.758 to 0.898) for DN, and 0.826 (95% CI, 0.783 to 0.866) for DPN. Conclusion: Exhibiting robust identification performance in both internal and independent external populations, these models hold potential as clinical tools for detecting microvascular complications in patients with T1D.