• The performance of Deep learning in outcome prediction of brain stroke
  • Faezeh pourshakori (corresponding author),1,* Mohammad Amoozadeh,2
    1. Student research committee, Anzali International Medical Campus, Guilan University of Medical Sciences, Guilan, Iran
    2. Student research committee, Anzali International Medical Campus, Guilan University of Medical Sciences, Guilan, Iran


  • Introduction: Stroke is a leading cause of mortality globally, with over 12 million incidents reported annually. It manifests in two primary forms: ischemic stroke, resulting from arterial blockage, and hemorrhagic stroke, caused by blood vessel leakage. Accurate prediction of stroke outcomes is crucial for guiding treatment decisions and minimizing disability, particularly as interventions like thrombolysis carry risks such as intracranial hemorrhage and gastrointestinal bleeding. Therefore, proper patient selection is essential. Traditional predictive methods, including Follow-up Infarct Volume (FIV), primarily rely on radiological assessments and are limited by their focus on specific clinical data subsets, often lacking comprehensive analysis. Recent advances in predictive systems have emerged, notably through the application of deep learning (DL), a significant branch of artificial intelligence. Studies indicate that DL can utilize imaging biomarkers for outcome prediction, leveraging architectures like deep neural networks and long short-term memory recurrent neural networks. These models are adept at analyzing complex, non-linear relationships between imaging and clinical data, making them particularly valuable in the context of stroke prediction. This review synthesizes recent research on DL approaches for stroke outcome prediction, examining their effectiveness compared to traditional methods and highlighting their potential to revolutionize clinical practice.
  • Methods: This review gathered recent studies from electronic databases such as PubMed, Scopus, and Google Scholar by utilizing the search terms "deep learning," "prediction," and "stroke." Following a thorough examination of the literature, a comprehensive conclusion was drawn.
  • Results: The review includes 10 studies employing various DL techniques, including DeepSM and DeepSurv, with models such as OEDL and SMOTEENN demonstrating considerable performance. The integration of clinical factors (ranging from 1 to 60) with imaging features notably enhanced the predictive accuracy of DL models. The outcomes were measured using the modified Rankin Scale (mRS) at three months post-stroke, revealing an area under the curve (AUC) was ranged from 0.779 to 0.830 and an improvement noticed from 0.779 to 0.88 when combining imaging and clinical data. However, specific AUC values for deep learning models focused solely on heart stroke prediction were not detailed in the provided search results. Sensitivity was approximately 0.79 while specificity was around 0.77. Further research may be needed to establish a comprehensive overview of AUC values across different deep learning.
  • Conclusion: In conclusion, the fusion of deep learning and clinical data significantly enhances the prediction of favorable reperfusion outcomes. The analysis of feature importance highlights both established and novel imaging characteristics with predictive significance. Furthermore, it is emphasized that outcome predictions should extend beyond infarct volume, as follow-up diffusion-weighted imaging provides additional prognostic insights, ultimately leading to better patient management and improved recovery trajectories. This innovative approach represents a promising advancement in stroke care, paving the way for more personalized and effective treatment strategies.
  • Keywords: deep learning; prediction; prognosis; outcome; stroke