EBM Model for Acromegaly Patient Outcomes
Methodology | Feature Importance | Interactive Tool
Feature Importance⌗
- EBM selected among cross-validated 1,200 experiments with different model families
- Extensive hyperparameter optimization
- Fitted to entire dataset
- Identifies 15 key features
- Used for feature importance analysis
Feature impact on log-odds scale from single EBM:
Open directly EBM features importances in a new tab.
Interactive Prediction Tool⌗
Clinical Prediction Tool: Quick Reference Guide⌗
| Step | Action | Notes |
|---|---|---|
| 1. Input Data | • Select features from dropdown • Enter corresponding values • Use 1/0 for binary features |
More data = better prediction |
| 2. Generate | Missing values filled with averages from centre data | |
| 3. Review Results | • Check risk level on gauge • Note confidence rating • View distribution of predictions |
“Uncertain” indicates model disagreement or closness to threshold |
Understanding the SHAP Visualization⌗
The SHAP waterfall plot represents a mathematical decomposition of the prediction into feature-wise contributions, based on cooperative game theory principles:
| Element | Interpretation | Clinical Relevance |
|---|---|---|
| Base Value | Expected value across population | Starting point before patient-specific factors |
| Feature Contributions | Marginal effects in log-odds space | How each clinical variable shifts the prediction |
| Directional Impact | Sign of the contribution (+ or -) | Blue increases risk, red decreases risk |
| Magnitude | Absolute SHAP value | Quantifies feature importance on logarithmic scale |
| Sequential Build | Additive composition property | Features ordered by absolute impact, with largest at chart bottom |
| Final Value | Sum of base value and all contributions | The model’s cumulative assessment |
Clinical Application Principles⌗
- The model operates in log-odds space, capturing non-linear relationships in clinical variables
- Focus on features with largest SHAP values—these drive prediction variance
- Model confidence reflects both ensemble consistency and distance from decision threshold; lower confidence occurs with high model variance or predictions near classification boundaries
- Consider counterfactual analysis: how would modifying controllable clinical factors affect the prediction?
- The tool provides decision support through feature attribution rather than causal inference