Deep Learning Outperforms ML for Stroke Risk in Carotid Plaque Patients

Overview

This study developed and compared deep learning (DL) and traditional machine learning (ML) models using carotid plaque ultrasound images to predict ischemic stroke risk. The ResNet-50 DL model achieved superior diagnostic accuracy (AUC 0.982) compared to ML models, demonstrating its potential as a precise, objective clinical tool for stroke risk stratification.

Background

Stroke is a leading cause of death and disability worldwide, with unstable carotid atherosclerotic plaques contributing to up to 25% of ischemic strokes. Early and accurate risk assessment of these plaques is critical for targeted prevention. Ultrasound is the primary screening modality but is limited by subjective interpretation and variability. While traditional ML models improve objectivity, they rely on manual feature extraction, which may miss subtle image characteristics. Deep learning, particularly convolutional neural networks, can automatically learn complex features from raw images, potentially enhancing predictive accuracy.

Data Highlights

Key Findings

• The ResNet-50 deep learning model achieved the highest diagnostic performance with an AUC of 0.982 for stroke risk prediction.
• ResNet-50 demonstrated superior accuracy (92.5%), sensitivity (96.4%), and specificity (89.7%) compared to traditional ML models.
• Among ML classifiers, logistic regression slightly outperformed support vector machine, but the difference was not statistically significant.
• The DL model improved AUC by 9.7% over the best ML model, indicating a clinically meaningful enhancement.
• Deep learning models can automatically extract complex subvisual features from ultrasound images, surpassing manual feature engineering limitations of ML.

Clinical Implications

The superior performance of the ResNet-50 deep learning model supports its use as a reliable, objective tool for early ischemic stroke risk stratification in patients with carotid plaques. Integrating such DL-based assessment into clinical workflows could enhance decision-making, enabling timely preventive interventions and potentially reducing stroke incidence. This approach may also reduce inter-observer variability inherent in conventional ultrasound interpretation.

Conclusion

This study validates that a lightweight deep learning model, specifically ResNet-50, significantly outperforms traditional machine learning classifiers in predicting stroke risk from carotid plaque ultrasound images, establishing it as a preferred clinical diagnostic tool for precise risk stratification.

References

1. Study Authors 2025 -- An Implementable Deep Learning Approach for Automated Risk Assessment of Stroke in Patients with Carotid Atherosclerotic Plaque