Development and external validation of an algorithm for self-identification of risk for microvascular complications in patients with type 1 diabetes.

Abstract

Microvascular complications, such as diabetic retinopathy (DR), diabetic nephropathy (DN) and diabetic peripheral neuropathy (DPN), are common and serious outcomes of inadequately managed type 1 diabetes (T1D). Timely detection and intervention in these complications are crucial for improving patient outcomes. This study aimed to develop and externally validate machine learning (ML) models for self-identification of microvascular complication risks in T1D population. Utilizing data from the Chinese Type 1 Diabetes Comprehensive Care Pathway program, 911 T1D patients and 15 patient self-reported variables were included. Combined with XGBoost algorithm and cross-validation, self-identification models were constructed with 5 variables selected by feature importance ranking. For external validation, an online survey was conducted within a nationwide T1D online community (N = 157). The area under the receiver-operating-characteristic curve (AUROC) was adopted as the main metric to evaluate the model performance. The SHapley Additive exPlanation was utilized for model interpretation. The prevalence rates of microvascular complications in the development set and external validation set were as follows: DR 7.0% and 12.7% (p = 0.013), DN 5.9% and 3.2% (p = 0.162) and DPN 10.5% and 20.4% (p < 0.001). The models demonstrated the AUROC values of 0.889 for DR, 0.844 for DN and 0.839 for DPN during internal validation. For external validation, the AUROC values achieved 0.762 for DR, 0.718 for DN and 0.721 for DPN. ML models, based on self-reported data, have the potential to serve as a self-identification tool, empowering T1D patients to understand their risks outside of hospital settings and encourage early engagement with healthcare services.