# Development of a machine learning-based radiomics model of perivascular adipose tissue for predicting stroke risk in patients with asymptomatic carotid stenosis: a multicenter study

## TL;DR

This study developed a machine learning model using CT scans of perivascular fat to predict stroke risk in patients with carotid artery narrowing, even when they show no symptoms.

## Contribution

The novel contribution is a combined radiomics and clinical model using perivascular adipose tissue features to predict stroke risk in asymptomatic carotid stenosis patients.

## Key findings

- Nine optimal radiomics features from perivascular adipose tissue were identified for stroke prediction.
- The XGBoost model achieved the highest AUC values across multiple datasets, showing strong predictive performance.
- The combined model incorporating clinical and radiomics features achieved AUCs up to 0.911, indicating high accuracy in predicting stroke risk.

## Abstract

Our work aims to develop and evaluate a combined model that integrates clinical features, conventional computed tomography angiography (CTA) features, and radiomics features of perivascular adipose tissue (PVAT) to identify asymptomatic carotid stenosis (ACS) patients at high risk for short-term stroke.

We enrolled 582 ACS patients confirmed by CTA from three medical centers and divided them into a training set (n = 188), an internal validation set (n = 85), and two independent external validation sets (set 1, n = 157; set 2, n = 152). Radiomics features of PVAT were extracted from CTA images, and dimensionality reduction was performed to identify predictive features. Five machine learning classifiers were employed to construct radiomics models, and the model with the highest AUC was selected to generate the radiomics score (Rad-score). Clinical factors associated with stroke were determined using univariate and multivariate logistic regression analyses to construct a clinical model. A combined model integrating clinical factors and the Rad-score was subsequently developed, and a nomogram was created to provide a visual tool for stroke risk prediction. We assessed model performance comprehensively through calibration curves, discrimination analysis, reclassification, and clinical application.

A total of nine optimal radiomics features were ultimately selected from the CTA images. Across the four datasets, the AUC values of the five classifier models ranged from 0.643 to 0.869, 0.716 to 0.826, 0.651 to 0.858, and 0.638 to 0.848, respectively, with the XGBoost model demonstrating the best performance. The combined model, incorporating hypertension, soft plaque, and the Rad-score as variables, achieved AUCs of 0.911, 0.868, 0.882, and 0.871, respectively, across the four datasets.

A combined model based on PVAT imaging features around carotid plaques can effectively predict the short-term stroke risk in ACS patients. This model may be expected to provide an important auxiliary tool for clinical prognosis assessment and treatment decisions, with potential clinical application value.