HEALTHCARE AI ANALYSIS
A Pilot Study on Prediction Models for Rehabilitation Outcomes in Ischemic Stroke Patients Based on Explainable Machine Learning
This study developed and externally validated an explainable machine learning model to predict rehabilitation outcomes for ischemic stroke patients. Using retrospective data from 200 patients, we selected 5 key predictors via the Boruta algorithm, capturing functional status, treatment intensity, inflammatory markers, and demographic factors. Among eight algorithms evaluated through a multi-criteria decision framework (integrating ROC, residual analysis, and stability metrics), logistic regression demonstrated the most robust and interpretable performance (test AUC=0.90, specificity=0.81), balancing discrimination with minimal overfitting. The model achieved strong external validation performance (AUC=0.98) with good calibration (Brier score=0.15) and positive clinical utility across threshold probabilities of 0.1-0.7, as confirmed by decision curve analysis. SHAP analysis revealed that post-treatment Barthel Index, pre-treatment functional status, and treatment duration were primary drivers of favorable outcomes, while age and white blood cell count negatively influenced prognosis. This parsimonious model effectively predicts outcomes using routine clinical data, balancing accuracy with transparency to offer a practical tool for personalizing rehabilitation strategies. Further validation in larger, multicenter cohorts is warranted.
Authors: Jingya Han, Qingyan Ma, Shiyu Meng, Yaorui Yang
Published: ICPHDS 2025, November 21-23, 2025, Yantai, China
Executive Impact Summary
Our analysis reveals how explainable machine learning can revolutionize stroke rehabilitation planning, leading to more personalized and effective patient care.
0.00
Predicted Outcome AUC
0
Key Predictors Identified
0.00
Clinical Utility Net Benefit
0.00
Brier Score (Calibration)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Methodology Flow
Model Performance Spotlight
Model Comparison
Individualized Prediction
Enterprise Process Flow: Model Development & Validation
Data Partitioning
→
Boruta Feature Selection
→
Model Development (8 ML Algorithms)
→
Multi-Criteria Model Selection
→
External Validation & Clinical Utility
0.98
External Validation AUC
The Logistic Regression model demonstrated robust discrimination, achieving a strong AUC of 0.98 on an independent external validation dataset, confirming its generalizability and consistent predictive accuracy across diverse patient populations.
| Feature | Logistic Regression (Selected Model) | Other ML Models (e.g., Random Forest, XGBoost) |
|---|---|---|
| Test AUC | 0.90 (Highest) | 0.59 - 0.86 |
| Train-Test AUC Gap | 0.06 (Smallest) | 0.12 - 0.26 (Higher, indicating overfitting) |
| Calibration (Brier Score) | 0.15 (Good Calibration) | Not explicitly quantified for others, but higher variability in residuals observed. |
| Interpretability | High (SHAP) | Lower (Often "black-box") |
| Overfitting Risk | Low | Higher, particularly for complex non-linear models. |
Logistic Regression demonstrated the best balance of discrimination and generalizability, outperforming more complex models that showed signs of overfitting, especially in the test set.
Individualized Prediction Example: Patient Prognosis
For a 79-year-old patient with BMIB = 40, BMIA = 60, NDT = 42, and WBC = 5.68x109/L, the model generated a final log-odds prediction of 2.655, translating to approximately 0.74 probability for a favorable outcome.
Key Influencing Factors:
- Age: -4 log-odds (Largest negative impact, increasing baseline risk due to advanced age).
- NDT (Number of Treatment Days): +3.21 log-odds (Strongest positive adjustment, reflecting benefit from extended rehabilitation).
- BMIB (Pre-treatment Barthel Index): +0.5 log-odds (Moderate positive contribution, higher initial functional status).
- WBC (White Blood Cell Count): -0.19 log-odds (Slight negative impact, elevated inflammation negatively affecting prognosis).
This demonstrates how the model integrates multi-dimensional clinical data into a coherent, quantifiable risk assessment, providing clinicians with transparent insights into individual patient trajectories.
Quantify Your AI Potential
Estimate the potential savings and reclaimed hours by integrating an explainable AI model into your stroke rehabilitation program.
Estimated Annual Savings
$0
Total Hours Reclaimed Annually
0
Your AI Implementation Roadmap
A phased approach to integrating explainable AI for stroke rehabilitation, ensuring seamless adoption and measurable impact.
Phase 1: Discovery & Data Integration
Assess existing clinical data infrastructure, identify relevant data sources (EMRs, lab systems), and establish secure data pipelines for initial model training and validation. Define key rehabilitation outcomes and success metrics.
Phase 2: Model Customization & Validation
Adapt the core AI model to your specific patient population and clinical workflows. Conduct rigorous internal and external validation using your institutional data, ensuring high accuracy, generalizability, and interpretability aligned with clinical guidelines.
Phase 3: Clinical Integration & Training
Integrate the validated AI model into your existing clinical decision support systems. Provide comprehensive training for clinicians and rehabilitation specialists on interpreting AI predictions, leveraging SHAP explanations, and incorporating insights into personalized patient care plans.
Phase 4: Monitoring & Continuous Improvement
Implement continuous monitoring of model performance, data drift, and clinical outcomes. Establish feedback loops with clinicians to refine the model, update features, and ensure ongoing relevance and ethical use, driving sustained improvements in rehabilitation strategies.
Ready to Transform Rehabilitation Outcomes with AI?
Book a complimentary 30-minute strategy session with our AI experts to explore how explainable machine learning can enhance patient prognosis and optimize resource allocation in your institution.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat