CardioFormer Cardiovascular EHR Dataset
Model-ready datasets for the CardioFormer cardiovascular EHR self-supervised learning project. Source code and experiment results: https://github.com/healme-225040511/CardioFormer
Contents
| Folder | Description |
|---|---|
cleaned_v2/ |
Deep-cleaned baseline master table, parsed troponin event series, audited patient splits, cleaning reports (missingness, zero-fill audit, category mappings) |
sequence_v1/ |
Model-ready SSL / downstream parquet splits (baseline + up to 16 troponin event slots), fitted preprocessors (categorical vocabs, continuous specs), smoke-test batch |
sequence_v1_strict_ssl/ |
Strict leakage-safe SSL sequence assets used for the revised experiment suite |
downstream_revised/ |
Task-specific downstream datasets (diagnosis multilabel, ordinal Killip/SCAI, peak-time bins) |
raw_data/ |
Raw source EHR exports (2026-02-11 completed versions): 3.8w_completed_20260211.xlsx (38,232 rows x 137 cols) and 2.7w_completed_20260211.xlsx (38,757 rows x 149 cols). Contains direct identifiers (patient name, registration number) - keep private |
downstream_prospective_peak_landmarks/ |
Prospective peak-landmark datasets: predicting remaining hours to a future troponin peak from fixed landmarks (first/second troponin, 6h, 12h) |
Splits
Patient-level, leakage-safe splits. Patients sharing a registration number across source datasets never cross split boundaries. The downstream pool (20% of grouped patients) was reserved before fitting any preprocessing statistics.
Privacy
The processed folders are derived from de-identified clinical EHR exports. raw_data/ contains direct patient identifiers and must never be made public. This repository is private;
do not redistribute patient-level data publicly.
Reports
See the GitHub repository for the full data-cleaning, training-data engineering, and pretraining reports (Markdown + standalone HTML).
- Downloads last month
- 7