CODE II dataset

A manifest.csv file (mapping each exam record to its corresponding package) and the original RECORDS.txt index are included to help locate specific exams within the packages. File naming convention Each tracing is stored as a pair of files following the pattern ecg_{exam_id}_reg{register_num}.hea / .dat (WFDB header and signal data files). Exams with multiple tracings produce multiple numbered pairs (e.g., ecg_533310_reg1, ecg_533310_reg2, ecg_533310_reg3).

Signal properties - Leads: 12 standard leads (DI, DII, DIII, AVR, AVL, AVF, V1–V6) - Unit: millivolts (mV) - Sampling rate: 300, 500, 600, or 1000 Hz (device-dependent; stored in each header file) - Duration: 7–12 seconds per tracing - Format: WFDB format 16 (16-bit integer) Leads DIII, AVR, AVL, and AVF are reconstructed from the 8 essential leads (DI, DII, V1–V6) using standard ECG equations. Signals are stored as recorded, after unit conversion (raw ADU → mV) and lead reconstruction; no additional preprocessing (filtering or resampling) was applied.

Researchers wishing to apply the preprocessing pipeline used in the accompanying manuscript should refer to: github.com/antonior92/ecg-preprocessing Reading the signals Signals can be read using the wfdb Python library. Example: import wfdb, then record = wfdb.rdrecord('ecg_533310_reg1') to load a record; record.p_signal returns the signal array in mV, and record.fs returns the sampling rate in Hz. A full description of the dataset, curation methodology, and the 66 CODE diagnostic classes is provided in the accompanying manuscript: Abreu, P. E. O. G. B. et al. CODE-II: A large-scale dataset for artificial intelligence in ECG analysis. npj Digital Medicine (2026). doi.org/10.1038/s41746-026-02704-4