Prediction of stroke events in patients with type 2 diabetes mellitus by interpretable machine learning based on...

Stroke is one of the leading causes of mortality, and patients with type 2 diabetes mellitus (T2DM) have a higher incidence of stroke. However, research on the imaging characteristics of plaques and perivascular adipose tissue (PVAT) in this patient population remains limited. This study therefore aimed to develop and validate a machine learning-based combined model to predict acute stroke events in patients with T2DM and assess its utility in stratifying patients into different risk categories based on follow-up outcomes.

In this multicenter study, a total of 494 computed tomography angiography (CTA) datasets from patients with T2DM were retrospectively collected from The Fifth Affiliated Hospital of Wenzhou Medical University, The Second Affiliated Hospital of Wenzhou Medical University, and Lishui People's Hospital and divided into four sets: training (n=193), internal testing (n=84), external validation 1 (n=105), and external validation 2 (n=102). Based on the magnetic resonance imaging findings, the patients were divided into a stroke group and a non-stroke group. PVAT features were extracted from CTA, and perivascular fat density (PFD) was determined. A combined model was developed by integrating radiomics scores with PFD and clinical factors via the extreme gradient boosting (XGBoost) algorithm. The model's prediction process was illustrated with the SHapley Additive exPlanation (SHAP) method, and its prognostic value was evaluated with Kaplan-Meier analysis.

In this study, 167 patients with T2DM (33.8%) who experienced ischemic stroke (IS) were classified into the stroke group, while 327 patients with T2DM (66.2%) were classified into the non-stroke group. Through application of variance thresholding, SelectKBest, and least absolute shrinkage and selection operator, seven radiomic features were ultimately selected from CTA images to construct the radiomics model. After univariate and multivariate logistic regression analysis, total cholesterol (P=0.033) and hypertension (P=0.028) were identified as independent risk factors for IS. The combined model demonstrated substantial accuracy and robustness, with an area under the receiver operating characteristic curve of 0.955, 0.847, 0.856, and 0.876 in the training, internal testing, external validation 1, and external validation 2 cohorts. SHAP analysis revealed that Exponential_glszm_SizeZoneNonUniformity and Wavelet-HLL_firstorder_Range were the most important features. Event-free survival (EFS) analysis demonstrated that the model could effectively determine patient prognosis. Results from univariate and multivariate Cox regression analyses identified the independent prognostic predictors of follow-up ischemic events to be stroke status [hazard ratio (HR) =3.916; 95% confidence interval (CI): 1.792-6.558; P<0.001] and predicted stroke status (HR =1.352; 95% CI: 1.317-4.777; P=0.030), indicating these factors are associated with the occurrence of ischemic cerebrovascular events during follow-up.

The combined XGBoost model incorporating PVAT features accurately predicted stroke events in patients with T2DM and provided risk stratification for patients.