Dataset DOI: https://doi.org/10.5061/dryad.fj6q57482

This Dryad deposit contains the de-identified TCGA-BRCA clinical annotation table (age provided as ranges), the Stata multiple imputation export, the modeling table used for the Python SHAP validation, the cohort gene expression matrix, a per-column data dictionary, the Stata scripts (published at Zenodo under the MIT license), and descriptive figures. The primary article is published in JAMIA Open, 2026;9(2):ooaf177, https://doi.org/10.1093/jamiaopen/ooaf177.

Primary article: Borges J. Auditing shortcut learning and misclassification in artificial intelligence-based breast cancer genomic subtyping. JAMIA Open. 2026;9(2):ooaf177. https://doi.org/10.1093/jamiaopen/ooaf177 (PMID 41937807, PMCID PMC13044511)

Description of the data and file structure

The study is a retrospective secondary analysis of The Cancer Genome Atlas Breast Invasive Carcinoma cohort (TCGA-BRCA) obtained from the UCSC Xena platform. A multinomial logistic regression model was trained on clinical predictors only (estrogen receptor status, progesterone receptor status, HER2 status, tumor stage, and age at diagnosis) to predict PAM50 intrinsic subtypes, and a pseudo-SHAP perturbation method implemented in Stata was used to quantify the marginal influence of each predictor and to detect shortcut learning. Canonical SHAP values were computed in Python (XGBoost and the shap package) on the same predictors for validation.

The deposit contains the de-identified clinical annotation table that the Stata pipeline starts from, the multiple imputation export produced by the pipeline, the six-column modeling table used by the Python notebook, the gene expression matrix of the cohort, a complete per-column data dictionary, the Stata scripts (published at Zenodo under the MIT license) and four descriptive figures (published at Zenodo under CC BY 4.0).

All clinical tables were de-identified for this deposit as described under Human subjects and de-identification. In particular, age at diagnosis is provided only as an age range in every file, so the deposited tables reproduce the study design but not the continuous age variable used in the published model; see Reproducibility notes.

Files and variables

TCGA_BRCA_clinicalMatrix_deidentified.tsv

One row per TCGA-BRCA sample (1,247 samples, 1,247 distinct sample barcodes) and 192 columns. The columns are the TCGA clinical data elements distributed by UCSC Xena for the cohort, in the original order, with two columns removed (days_to_birth, form_completion_date) and four columns transformed (two age columns binned, two gender columns recoded). Column families:

Variables used by the published model: PAM50Call_RNAseq (outcome), ER_Status_nature2012, PR_Status_nature2012, HER2_Final_Status_nature2012, AJCC_Stage_nature2012 (grouped I, II, III) and Age_at_Initial_Pathologic_Diagnosis_nature2012 (continuous in the analysis, age range in this deposit).

De-identified columns in this file: age_at_initial_pathologic_diagnosis and Age_at_Initial_Pathologic_Diagnosis_nature2012 hold age ranges (<30, 30-39, 40-49, 50-59, 60-69, 70-79, 80+); gender and Gender_nature2012 hold numeric codes. Every column is described individually in data_dictionary_v2.csv.

TCGA_BRCA_clinicalMatrix_deidentified.dta

The same 1,247 x 192 table saved from Stata 18 in Stata 13 file format (release 117) with saveold. Stata limits variable names to 32 characters, so long TCGA names are truncated and 16 names that collided after truncation were replaced by Stata with positional names (v48, v58, v59, v115, v116, v118, v119, v121, v122, v126, v127, v141, v174, v187, v190, v192); for each of these the original TCGA column name is stored in the Stata variable label and listed in the original_name column of data_dictionary_v2.csv (for example v48 = age_at_initial_pathologic_diagnosis). Numeric variables use Stata system missing (.) for missing values; string variables use the text NA. The age variables are strings holding age ranges and the gender variables are numeric codes.

Opening the file: Stata 13 or later (use), R haven::read_dta() ReadStat-based; the older foreign package cannot read any file format newer than Stata 12 and is not expected to work), Python pandas.read_stata() or pyreadstat.read_dta(). The file was verified to open with pyreadstat 1.3.6 (ReadStat), the same C library used by R haven and by JASP.

capstone_model_data_for_shap_export_deidentified.csv

The dataset exported from Stata after steps 1 to 5 of the pipeline, in Stata multiple imputation long format (mi set mlong). It has 9,347 rows and 209 columns:

The column to original name mapping for all 209 columns is given in data_dictionary_v2.csv (column original_name). De-identified columns: age_at_initia_o and v48 hold age ranges (imputed non-integer ages were floored before binning); gender and gender_nature2012 hold numeric codes; days_to_birth and form_completion_date were removed.

AI-BIAS_cleaned_TCGA_PAM50_model_dataset_deidentified.csv

The six-column table was loaded by the Colab notebook PAM50_SHAP_Colab_Notebook.ipynb (GitHub repository) to train the XGBoost classifier and compute canonical SHAP values (SHAP value array shape 1,231 x 5 features x 5 classes).

The table has 1,231 rows and no missing values. It was produced from the same TCGA-BRCA cohort during the capstone work; 767 rows are exact duplicates of another row in the original export and were retained unchanged (apart from age binning) so that the deposited table matches the notebook input.

TCGA.BRCA.sampleMap_HiSeqV2.tsv

The IlluminaHiSeq gene expression matrix of the cohort as distributed by UCSC Xena (dataset TCGA.BRCA.sampleMap/HiSeqV2): 20,530 rows (genes; first column sample holds the gene symbol) and 1,218 sample columns headed by TCGA sample barcodes; values are log2(RSEM normalized count + 1). The file is unchanged from the Xena distribution (SHA-256 9430ba08c8f3ad70f6e61cc75dc36f1f34cce1c50b6d707bf63825512f32d78a) and contains no clinical or demographic fields. It is included as the genomic reference for the cohort; the published model uses clinical predictors only, so it is not an input to the scripts.

data_dictionary_v2.csv

600 rows, one per column of every tabular file (including the .dta and the expression matrix), with: file, column_name, original_name (the TCGA or Xena name when a Stata name differs), description, n_non_missing, n_distinct, example_values (up to five observed values).

Missing data

Every empty cell in the tab delimited and CSV files was replaced by the text NA (1,247 x 192 clinical matrix: 116,419 NA cells; multiple imputation export: 1,138,576 NA cells, most of them in the imputed rows, which carry values only for the registered variables and identifiers; six column modeling table: none). In the Stata file, numeric variables use Stata system missing (.) and string variables use NA. Missingness reflects the TCGA source annotations: many _nature2012 columns are populated only for the 952 samples included in the 2012 TCGA Network analysis, and PAM50Call_RNAseq is available for 956 samples. The 691 complete cases analysed in the article are the samples with a PAM50 call andnon-missingg ER, PR, HER2, stag,e and age.

Human subjects and de-identification

The data derive from TCGA-BRCA, a public, consented, de-identified research resource accessed through UCSC Xena; no direct identifiers exist in the source. Following Dryad's Human Subjects and Identifiable Information policy, the following additional de-identification was applied uniformly to every clinical table in this deposit (rules R1 to R8):

TCGA sample and patient barcodes are public study identifiers and were retained. The expression matrix contains no demographic fields. Descriptive figures are aggregate plots. No file in this deposit contains an exact age, a date of birth, a calendar date of any event, or any direct identifier.

Code and software

Reproducibility notes

1. The scripts were written against the continuous age variable. Because the deposited tables carry age ranges (Dryad policy for human subjects data), rerunning the scripts on the deposited files will not reproduce the continuous age term of the published model. To reproduce the published estimates exactly, download the public clinical matrix TCGA.BRCA.sampleMap/BRCA_clinicalMatrix from UCSC Xena (link under Access information), import it into Stata (import delimited) and run the scripts; the deposited de-identified table is otherwise identical to that source apart from rules R1 to R8.
2. capstone_step_1_to_3_data_cleaning.do begins with use "your_datafile.dta"; replace the placeholder with the path of TCGA_BRCA_clinicalMatrix_deidentified.dta or of the Stata import of the Xena matrix. The scripts refer to the age and last contact variables under the names age_at_initia_o and days_to_date_t (renamed from age_at_initial_pathologic_diagno and days_to_date_of_last_contact_nat, respectively).
3. In the deposited Stata f,ile the gender variables are already numeric, so the encode gender_nature2012, gen(gender_num) line of step 4 is not applicable (the variable is dropped again in step 5 and does not enter the model).
4. Multiple imputation uses random draws; set a seed (set seed) before mi impute for bitwise reproducibility of the imputed rows.

Access information

Data were derived from the following public sources:

Other locations of related material:

Licensing

The author confirms that publication of the data files under CC0 is intended and accepted, and that the MIT license of the code in the GitHub repository is not affected by it.

Contact

Julian Yin Vieira Borges, MD, MS. Department of Computer Science, Boston University Metropolitan College, Boston, MA, United States. jyborges@bu.edu. ORCID https://orcid.org/0009-0001-9929-3135

Human subjects data

Human subjects de-identification statement. All data are derived from The Cancer Genome Atlas Breast Invasive Carcinoma cohort (TCGA-BRCA), a public research resource accessed through the UCSC Xena platform. TCGA participants provided informed consent for the collection and open sharing of de-identified clinical and molecular data under IRB-approved protocols administered by the NCI/NHGRI TCGA program, and the data are distributed by the NCI Genomic Data Commons and UCSC Xena without direct identifiers. No new data were collected from human participants, and no identifiable private information was accessed; the study is exempt under 45 CFR 46.101(b)(4). For this Dryad deposit the following additional de-identification was applied uniformly to every clinical table: exact age at diagnosis was replaced by age ranges (<30, 30-39, 40-49, 50-59, 60-69, 70-79, 80+, so that ages of 90 and above are aggregated); the field days_to_birth (exact age in days) and the field form_completion_date (a calendar date) were removed; gender was recoded to numeric codes whose definitions are not published; empty cells were set to NA. Retained fields are TCGA sample and patient barcodes (public study identifiers), year of diagnosis (year only), and interval fields expressed in days from diagnosis, which are not calendar dates. The gene expression matrix contains no demographic fields, and the deposited figures are aggregate plots. No file in this deposit contains an exact age, a date of birth, a calendar date of any event, a name, an address, or any other direct or indirect identifier. The full rule set (R1 to R8) and the per-column data dictionary are provided in the README.