data/README.md

Data

predictions/

One folder per reader, then per training direction and seed:

• deployed-protocol_clam-sb_recipe-A/ — the deployed reader's protocol: CLAM-SB on 400 tiles at 0.5 um/px, our deployed encoder. t2y: trained on TCGA-BRCA (490 slides), external test Yale-HER2 (192). y2t: trained on Yale-HER2, external test TCGA-BRCA. Seeds 1-10; the deployed reader is t2y/seed4.
• r11o_published-protocol/ — the published reader's protocol with our deployed encoder: attention MIL with non-gated attention on every tissue tile at 1.0 um/px, trained on TCGA-BRCA (917) + CPTAC-BRCA (382), external test Yale-HER2. Seeds 1-8 (40 models).
• r12a_published-protocol/, r12b_published-protocol/ — the same protocol with the R12 encoders.

Files in each seed folder:

• predictions.csv — training cohort, 5-fold cross-validation: fold, split (train/val/test), slide_id, case_id, label, score. Rows with split == test are the out-of-fold predictions.
• external_predictions.csv — every fold model's score on every external slide: fold, split (external), slide_id, case_id, label, score. The seed's reader is the mean over its five folds.
• results.csv — per-fold, pooled and external-ensemble metrics (AUC with DeLong and bootstrap intervals, threshold, sensitivity, specificity, PPV, NPV).
• folds.csv — which fold each slide's case was dealt into (case-stratified, fold seed 1337).

label is HER2 status (1 positive, 0 negative): for TCGA-BRCA from GDC clinical data, for CPTAC-BRCA from its proteogenomic ERBB2 status, for Yale-HER2 from the collection's case labels. score is the probability of HER2-positive.

paper/

paper-data/ holds the data files our paper's build reads (tables, figure data, macros); book-data/ holds the files it reads from our technical report's data folder.