Predicting the risks for stroke, cardiovascular disease, and peripheral vascular disease among people with type 2 diabetes with artificial intelligence models: a systematic review and meta-analysis

Abstract

Objectives: This systematic review and meta-analysis aim to explore the performance of machine learning algorithms in predicting the risk of macrovascular complications among individuals with T2DM, specifically, the predictive capabilities of AI models in forecasting stroke, CVD, and PVD in LMICs. Design: Systematic review and meta-analysis of studies reporting on AI prediction models for macrovascular complications in T2DM patients. Setting: The review included studies conducted in various healthcare settings, primarily from LMICs, upper-middle-income countries (UMICs), and high-income countries (HICs). Participants: 46 studies were included, with a total of 184 AI models. Participants were diverse in age, sex, and geographical locations, reflecting a broad range of healthcare settings. Interventions: The intervention analyzed was the application of AI models, including machine learning algorithms, to predict macrovascular complications such as stroke, CVD, and PVD. Primary and Secondary Outcome Measures: The primary outcome was the predictive performance of AI models, measured by the area under the receiver operating characteristic curve (AUROC). Secondary outcomes included subgroup analyses based on predictor types and an assessment of AI model applicability in low-resource settings. Results: Twelve included studies yielded 184 AI models with an overall AUROC of 0.753 (95%CI: 0.74-0.766; I2=99.99%; p<0.001). For 80 models of cardiovascular outcomes, an AUROC of 0.741 (95%CI: 0.721-0.76; I2=99.78%; p<0.001) was obtained. Meanwhile, 25 models of peripheral vascular disease and 38 models of cerebrovascular diseases obtained AUROCs of 0.794 (95%CI: 0.758-0.831; I2=97.23%; p<0.001) and 0.77 (95%CI: 0.743-0.797; I2=99.73%; p<0.001) respectively. Subgroup analysis revealed that models with lab-only predictors were superior to those with mixed or no-lab predictors. This signalled the lack of AI capability for history-taking and physical examination data alone, primarily available in low-resource settings. Conclusions: Artificial intelligence is promising in predicting diabetes complications. Nevertheless, future studies should explore accessible features in low-resource settings and employ external validation to ensure the robustness of the prediction models.