Clinical Text Imbalance Benchmark—Results (336 configurations), v2.
Abstract
This dataset contains per-configuration test metrics for a large-scale benchmark of clinical text classification under extreme class imbalance (49,035 French breast radiology reports; minority prevalence ≈0.33%). A factorial design varied two vectorisers (BoW, TF–IDF), 12 resampling methods plus a baseline, and 15 classifiers. The file ml_experiment_results.csv reports one row per executed configuration (n=336) with: Vectorizer, Sampler, Classifier, Accuracy, Balanced_Accuracy, ROC_AUC, PR_AUC, Precision_male, Recall_male, F1_male, Precision_female, Recall_female, F1_female, F1_macro, F1_weighted, TP, FP, TN, FN.
Full text
Supplementary Figures — Clinical Text Imbalance Benchmark (v2) Figure S1: PCA (BoW). Two-component PCA of BoW features. Blue = majority class; red = minority class. Points show high-dimensional sparsity with broad overlap and no linear separation. 1
Figure S2: PCA (TF–IDF). Two-component PCA of TF–IDF features. Blue = majority; red = minority. Relative scaling tightens variance but still shows strong overlap and weak linear separability. 2