What question did this study set out to answer?

May 4, 2026Diagnostics0 citationsOpen Access

Integrative Multidimensional Machine Learning Models for Stroke Prognosis: Age-Stratified and History Engineered Perspectives

GLGawon LeeHallym University SKSunyoung KwonHallym University SSSeung-Ho ShinSacred Heart Hospital

Key Points

The study aims to develop machine learning models for predicting mortality in stroke patients by integrating clinical and demographic factors.
Retrospective analysis of data from 1780 stroke patients admitted between 2018 and 2023.
Utilized Random Forest models to predict mortality at 1, 2, and 3 years post-admission.
Assessed model performance using AUC and SHAP values.
The model for patients under 60 achieved an AUC of 0.995 (95% CI: 0.98–1).
Pulse rate was consistently identified as the most important predictor of mortality risk.
Subgroup analyses based on age and medical history improved model interpretability and predictive power.

Abstract

Introduction: Stroke remains a leading cause of mortality and long-term disability worldwide. Accurate prognosis prediction is essential for timely intervention and personalized treatment planning. However, previous studies have often overlooked the role of patients’ medical history, age-specific risk factors, and time-dependent mortality patterns. This study aimed to develop and evaluate machine learning models for predicting mortality in stroke patients by incorporating vital signs, blood test results, demographic characteristics, and medical history, while also exploring subgroup-specific factors. Methods: We retrospectively analyzed data from 1780 stroke patients admitted to Hallym University Sacred Heart Hospital between 2018 and 2023. Input features included both original and binarized forms of vital signs and blood test values, along with age and medical history. Random Forest models were developed to predict mortality at 1, 2, and 3 years post-admission, as well as overall mortality. Model performance was assessed using AUC and 95% confidence intervals, and variable importance was evaluated using Mean Decrease Gini and SHAP values. Results: The highest predictive performance was observed in a model for patients under 60 using binarized input features, achieving an AUC of 0.995 (CI: 0.98–1). Across all models, pulse rate consistently emerged as the most important predictor. Additional key features included platelet count and diastolic blood pressure. SHAP analysis revealed that pulse rate was associated with higher mortality risk. Subgroup analyses based on age and medical history improved interpretability and predictive power. Conclusions: This study demonstrates that integrating clinical indicators with demographic and medical history variables can significantly enhance the accuracy and interpretability of mortality prediction models in stroke patients. The results underscore the importance of stratified modeling and continuous monitoring of vital signs, particularly pulse rate, to support precision stroke care.

问 AI

Bookmark

View Full Paper