همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Ensemble Machine Learning for Predicting Stroke Patient Survival

نویسندگان :

Seyedeh Maryam Mousavi¹ Samira Ahmadi² Solmaz Norouzi³

1- دانشگاه علوم پزشکی زنجان 2- دانشگاه علوم پزشکی زنجان 3- دانشگاه علوم پزشکی زنجان

کلمات کلیدی :

Survival Support Vector Machine،Random Survival Forest،Survival Support Vector Machine and Random Survival Forest Ensemble،Survival Analysis،Stroke

چکیده :

Abstract Background: Stroke is a leading global cause of death and disability, making accurate survival prediction crucial for clinical decision-making. Recent advances in machine learning provide robust tools for survival analysis, effectively handling complex, non-linear, and high-dimensional data. This study evaluated and compared the predictive performance of Survival Support Vector Machine (SSVM), Random Survival Forest (RSF), and their ensemble models in identifying factors influencing stroke patient survival. Methods: We evaluated the performance of Survival Support Vector Machine (SSVM), Random Survival Forest (RSF), and their ensemble models in predicting survival outcomes in a cohort of 332 stroke patients from Imam Khomeini Hospital in Ardabil, who were monitored for up to 15 years following their stroke (2007-2022). Collected variables encompassed demographic, clinical, and lifestyle factors, including age, sex, Education, comorbidities (e.g., hypertension, heart disease), stroke type (ischemic or hemorrhagic), and smoking history. Alongside the Cox model, we implemented SSVM and RSF, then constructed four ensemble models (Voting, Stacking, Bagging, and Boosting) by integrating their outputs to capitalize on the strengths of both algorithms. Performance was assessed using accuracy, sensitivity, specificity, area under the curve (AUC), and survival function estimation error. Results: Ensemble models outperformed the individual RSF, SSVM, and Cox models in predicting long-term mortality. The Voting ensemble achieved the highest performance, with an AUC of 0.855, overall accuracy of 0.833, 95% confidence interval of 0.747–0.962, sensitivity of 0.867, and specificity of 0.762. It demonstrated clear superiority over SSVM (AUC = 0.711), RSF (AUC = 0.828), and Cox (AUC = 0.768). Variable importance analysis in the optimal Voting model ranked age, hypertension, heart disease history, and hemorrhagic stroke type as the top predictors of long-term mortality. Conclusion: Integrating SSVM and RSF within ensemble frameworks enhances both predictive accuracy and interpretability in stroke survival analysis. The Voting ensemble, as the most accurate model, offers a valuable clinical tool for identifying high-risk patients and guiding targeted interventions.