Fig. 4.

Classification of patient response in non-pCR and pCR using ULM features. A Classification workflow: Model performance was evaluated based on LOOCV. Therefore, patients were split into a training group (n = 19) and a validation group (n = 1). Feature selection and classification via logistic regression were performed on the training group. The model was then used to predict the remaining patient. This procedure was repeated 20 times, after which model performance was evaluated based on predictive accuracy and AUC. The model design was optimised via a grid search, and the best predictive performances were reported. B Classification results for individual measurements, changes between measurements, and the combination of several measurements. C Features with selection frequency ≥50% among the different measurements. The selection frequency of a feature was determined as the percentage of times it was selected for classification among the 20 folds of LOOCV.