Predicting motor recovery after spinal cord injury using a stacked ensemble

Overview

This project develops a supervised machine learning model to predict 12-month motor recovery outcomes for individuals with spinal cord injury (SCI). Using data collected during the first 30 days post-injury, the model estimates the International Standards for Neurological Classification of Spinal Cord Injury (ISNCSCI) total motor score at 12 months. View the full rendered project page here

Objective

Predict the ISNCSCI total motor score (0-100) at 12 months post-injury using early clinical data, leveraging stacked ensemble learning for optimal predictive performance.

Dataset

• Source: Simulated dataset modeled after the European Multicenter Study about Spinal Cord Injury (EMSCI).
• Sample Size: 1,500 patients.
• Design: Includes realistic heterogeneity, variance, and ceiling effects.
• Input Features (Day 0-30):
  • Age
  • Sex
  • Cause of Injury
  • Neurological Level of Injury (NLI)
  • Baseline AIS Grade
  • Lower Extremity Motor Scores (LEMS)
  • Sensory Scores (Light Touch, Pin Prick)
• Target Variable: ISNCSCI Total Motor Score at 12 months (0-100).

Methodology

Modeling Approach: Stacked Generalization

The project uses stacked ensemble learning to combine diverse algorithms for improved predictive performance.

Level 0: Base Learners

• XGBoost (Gradient Boosting)
• SVM (RBF kernel) via kernlab
• MARS via earth
• GLMNet (Elastic Net / Lasso)

Level 1: Meta-Learner

• LASSO Regression to aggregate Level 0 predictions.

Validation Strategy

• 5-fold Cross-Validation across all base learners and the meta-learner.

Results

• R-Squared: ~0.74
• RMSE: ~14.8
• MAE: ~11.3

Subgroup Performance

• Paraplegic: R² = 0.65
• Tetraplegic: R² = 0.78
• AIS A: R² = 0.45 (Most difficult subgroup to predict)
• AIS D: R² = 0.75

Calibration

Overall calibration is strong, with slight underprediction in high recovery ranges (76-100).

Interpretability

Model interpretability was assessed using permutation-based and SHAP-based approaches.

Key Predictors

• Baseline Total Motor Score (strongest predictor)
• AIS Grade
• Lower Extremity Motor Score (LEMS)

Tech Stack

• Language: R (v4.5.1)
• Framework: tidymodels
• Key Libraries:
  • stacks
  • xgboost
  • ranger
  • earth
  • kernlab
  • DALEX
  • fastshap
  • tidyverse

Contact

For questions or collaborations, please contact me. Found an error? Open a ticket!