Imaging vs. Clinical Value Explorer
Predict pathologic tumour size from clinical variables, imaging features, or both, and measure the incremental value with leave-one-out RMSE.
- incremental value
- radiomics
- LOO cross-validation
- regression
Leave-one-out RMSE comparison
| Model | LOO RMSE (cm) | vs. clinical (%) | n |
|---|---|---|---|
| clinical only | 0.912 | 0.0 | 80 |
| imaging only | 0.544 | -40.3 | 80 |
| clinical + imaging | 0.545 | -40.2 | 80 |
Predicted vs. actual (combined model)
Learning curve
ClinicalImagingCombined
Incremental value must be measured out of sample
Adding imaging features always lowers in-sample residual error; only the leave-one-out RMSE can tell you whether it lowers *prediction* error. Shrink n and the combined model — with the most parameters — is the first to degrade, which is exactly the variance cost of extra predictors in a small cohort.