# Integrating cerebrovascular morphology and radiomics features for predicting stroke prognosis: a retrospective study

## TL;DR

This study develops a machine learning model that combines brain imaging features to predict stroke recovery outcomes, improving accuracy and reducing manual evaluation.

## Contribution

The novel framework integrates cerebrovascular morphology and radiomics features from multiple regions to predict stroke prognosis more accurately.

## Key findings

- The model achieved high accuracy (0.828) and AUC (0.942) in predicting 90-day stroke outcomes.
- Combining vascular morphology and radiomics features improved prediction performance significantly.
- 26 selected features from 1,066 extracted features were used to build the final classification model.

## Abstract

Accurately predicting 90-day Modified Rankin Scale (mRS) scores for acute ischemic stroke (AIS) patients is crucial for guiding treatment strategies. However, many existing mRS prediction methods rely on clinicians to manually evaluate relevant features, and the accuracy of feature quantification and model reproducibility still need to be further improved. This study proposes a machine learning framework that combines multimodal imaging features in order to predict 90-day mRS outcomes. A retrospective analysis was conducted on 86 AIS cases. Morphological features of the intracranial arterial and venous system were extracted from computed tomography angiography (CTA) images. Additionally, radiomics features were obtained from the ischemic lesion on diffusion-weighted imaging (DWI). Recognizing the significance of the peri-infarct penumbra in stroke prognosis, radiomics features were also extracted from the annular region surrounding the ischemic lesion. Redundant features were eliminated using a sparse representation method, and a sparse representation-based classifier was developed to predict mRS outcomes. Model performance was validated using cross-validation and independent test. A total of 1,066 features, including 40 vascular morphological features and 1,026 radiomics features, were extracted. Both feature types demonstrated statistical significance (P < 0.05). Ultimately, 26 features were selected to construct the classification model. The proposed model achieved robust performance on the independent test set, with a classification accuracy of 0.828, an area under the curve (AUC) of 0.942, sensitivity of 0.789, specificity of 0.900, positive predictive value of 0.937, and negative predictive value of 0.692. By integrating vascular morphological features with radiomics features from the ischemic lesion and peri-ischemic lesion regions in DWI, the proposed machine learning model provides accurate predictions of 90-day clinical outcomes for AIS, offering valuable insights for personalized stroke management.