Dynamic, Interpretable, Machine Learning-Based Outcome Prediction as a New Emerging Opportunity in Acute Ischemic
Ivan Petrović1, Sava Njegovan2, Olivera Tomašević2
1Faculty of Medicine, University of Novi Sad, Novi Sad, Serbia.
Stroke Research and Treatment
|April 2, 2025
Summary
This study developed interpretable machine learning models to predict stroke recovery outcomes, outperforming existing scores. Key predictors like age and treatment time offer insights for better patient care.
Area of Science:
- Neurology
- Artificial Intelligence
- Biomedical Informatics
Background:
- Machine learning (ML) models in clinical settings are often "black boxes," limiting trust and application.
- Stroke patient recovery is dynamic, making traditional predictive models less reliable over time.
- Interpretable ML offers a path to understanding predictions and improving patient care.
Purpose of the Study:
- To build and evaluate an interpretable ML model for predicting acute ischemic stroke (AIS) patient outcomes.
- To assess model performance at various recovery time points (baseline, 2-h, 24-h, discharge).
- To enhance the transparency and reliability of stroke outcome prediction.
Main Methods:
- Retrospective analysis of 355 AIS patients treated with alteplase.
- Utilized baseline, 2-h, 24-h, and discharge data as model inputs.
- Employed Support Vector Machine (SVM), Logistic Regression (LR), and Random Forest (RF) classifiers with SHAP and LIME for interpretability.
Main Results:
- Interpretable ML models demonstrated strong predictive performance (AUC 0.80-0.96), surpassing the DRAGON score (AUC 0.760).
- SHAP and LIME identified key predictors, including age, onset-to-treatment time, platelet count, NIHSS, and blood pressure.
- Models provided clear insights into factors influencing patient outcomes at different time points.
Conclusions:
- Interpretable ML models are effective for predicting functional outcomes in AIS patients treated with alteplase.
- Dynamic prediction using interpretable models aids in identifying modifiable factors for improved patient recovery.
- This approach enhances clinical decision-making by providing understandable and secure output.
