Development and Clinical Validation of a Stroke Risk Decision System for Patients With Carotid Artery Plaque Based on
AI
‐Powered Handheld Ultrasound and Laboratory Markers
Zhao Jinyi, Liao Mingjiao, Peng Yi, Yang Rui, Zhang Wenjun ABSTRACT
Objective
This study aimed to develop and validate a fusion model combining portable artificial intelligence (AI) ultrasound imaging with multidimensional laboratory data to stratify risk and support decision‐making in patients with carotid plaque‐related anterior circulation stroke, and to assess its predictive performance and clinical utility.
Methods
This prospective study enrolled 639 patients with carotid plaque via AI‐powered ultrasound at community health centers. Anterior circulation stroke (ACS)/transient ischemic attack (TIA) patients were excluded. All participants completed the 180‐day follow‐up. The primary outcome was new‐onset anterior circulation stroke or TIA, defined as focal neurological deficit > 24 h with CT/MRI‐confirmed infarction, or transient (< 24 h) deficit without evidence of acute infarction on neuroimaging. Outcomes were adjudicated by two neurologists (attending or above), with a third for disagreements. Participants were classified into ACS group ( n = 140) and non‐ACS group ( n = 499). Ultrasound features of carotid plaques and clinical laboratory data were compared between the two groups. Multivariable logistic regression was performed to identify factors associated with ACS, and a nomogram was constructed based on the regression model. Internal validation was conducted using the bootstrap method. Model performance was assessed in terms of discrimination (C‐index and area under the curve [AUC]), calibration (calibration curves), and clinical utility (decision curve analysis [DCA] and clinical impact curves [CIC]).
Results
(1) Significant differences were observed between the two groups in history of hypertension, smoking, hyperlipidemia, stroke, and myocardial infarction (MI), as well as plasma triglyceride (TG), low‐density lipoprotein (LDL), high‐density lipoprotein (HDL) levels, carotid plaque thickness, plaque area, plaque echogenicity percentage, degree of carotid stenosis, plaque continuity score, and plaque mutability score ( p < 0.05). No significant differences were found in sex, history of diabetes, total cholesterol (TC), plaque location, or plaque fibrous cap continuity analysis score ( p > 0.05) (2) Multivariable logistic regression identified history of stroke, history of myocardial infarction, plaque thickness, LDL, and HDL as independent predictors of stroke. A nomogram based on these factors demonstrated good discrimination (C‐index/AUC = 0.884), with a sensitivity of 84.29% and specificity of 72.55%, along with excellent calibration (Hosmer–Lemeshow test p = 0.250). Decision curve analysis showed superior net benefit across a range of risk thresholds compared with the null model, supporting a multi‐threshold stratified intervention strategy. Notably, when the predicted risk exceeded 75%, the model's predictions closely aligned with observed outcomes.
Conclusion
The stroke risk decision system developed using AI‐powered handheld ultrasound and laboratory indicators demonstrates clinical applicability for early prediction of anterior circulation ischemic stroke and is particularly suitable for early screening and risk stratification of high‐risk stroke populations in resource‐limited settings.