The Comparison and Interpretation of Machine-Learning Models in Post-Stroke Functional Outcome Prediction
Shih-Chieh Chang1, Chan-Lin Chu2,3, Chih-Kuang Chen2,4
1Department of Physical Medicine and Rehabilitation, Chang Gung Memorial Hospital at Linkou, Taoyuan 333, Taiwan.
Diagnostics (Basel, Switzerland)
|October 23, 2021
Summary
Predicting stroke recovery is possible using machine learning models and patient data collected at admission. Key predictors like balance and cognitive tests significantly influence functional outcomes, aiding resource allocation.
Area of Science:
- Neurology
- Medical Informatics
- Rehabilitation Medicine
Background:
- Accurate prediction of post-stroke functional outcomes is essential for effective medical resource allocation.
- The Post-Acute Care-Cerebrovascular Disease (PAC-CVD) program enrolled 577 patients, collecting 77 potential predictors at admission.
Purpose of the Study:
- To evaluate the efficacy of various machine learning (ML) methods in predicting functional outcomes (Barthel Index >60) at discharge.
- To identify key predictors influencing stroke recovery and understand their predictive value ranges.
Main Methods:
- Eight machine learning models were applied, with results integrated using a stacking method.
- Feature importance analysis, Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) plots were used to analyze predictor influence.
Main Results:
- The ML models achieved an Area Under the Curve (AUC) between 0.83 and 0.887, with Random Forest, stacking, logistic regression, and support vector machines showing high performance.
- Initial Berg Balance Test (BBS-I), initial Barthel Index (BI-I), and initial Concise Chinese Aphasia Test (CCAT-I) were the most significant predictors.
- Predictive power was strongest within specific ranges for BBS-I (<40) and BI-I (<60).
Conclusions:
- Machine learning models can effectively predict discharge Barthel Index scores using admission data.
- Understanding predictor value ranges through PDP and ICE plots enhances the interpretability and clinical utility of predictive models for stroke rehabilitation.
