Abstract WP296: Retrospective Analysis of Deep Learning Models for Acute Ischemic Stroke Prediction in Stroke Event Risk Assessment: The RADAIS Study

Background: The application of predictive analytics, particularly through artificial intelligence (AI) and machine learning (ML), has the potential to transform risk assessment in Acute Ischemic Stroke (AIS). This study aims to compare the performance of deep learning (DL) models against traditional machine learning (ML) algorithms in predicting AIS risk, using clinical data from patients with a history of cerebrovascular stroke events. Methods: In this retrospective analysis, we examined data from 5120 patients with documented AIS events, sourced from Rhythm Heart and Critical Care. Key predictive variables included hypertension, history of heart disease, occupation, average glucose levels, body mass index (BMI), and smoking status. To ensure unbiased model development, stroke history was blinded during training. We assessed the performance of various DL architectures. These were benchmarked against established traditional models, such as Logistic Regression (LR), AdaBoost, and Gaussian Naïve Bayes (GaussianNB). Model efficacy was evaluated using key metrics including accuracy, precision, prediction rate, and area under the receiver operating characteristic curve (AUC). Results: The RADAIS model, incorporating deep learning techniques, successfully generated stroke predictions for 5017 patients (97% of the cohort). The model demonstrated superior performance with an accuracy of 0.95, precision of 0.99, prediction rate of 0.97, and AUC of 0.97. In contrast, traditional machine learning models exhibited lower performance, with Logistic Regression (accuracy: 0.71, precision: 0.71, prediction: 0.69, AUC: 0.79), Gaussian Naïve Bayes (accuracy: 0.70, precision: 0.68, prediction: 0.73, AUC: 0.81), and AdaBoost (accuracy: 0.78, precision: 0.75, prediction: 0.82, AUC: 0.76). These results underscore the superior accuracy and precision of deep learning models in AIS prediction. Conclusion: The findings from the RADAIS study highlight the significant advantages of deep learning models over traditional machine learning methods for predicting the risk of Acute Ischemic Stroke. With a remarkable accuracy of 0.95 and an AUC of 0.97, deep learning-based models offer a more precise, reliable, and clinically relevant tool for stroke risk assessment, ultimately facilitating more informed decision-making in stroke management.