Processed dual-channel PCG–ECG window dataset and screening documentation for PhysioNet/CinC 2016 training subset-a

This deposit provides a self-contained data package derived from the PhysioNet/CinC 2016 Challenge training subset-a. It includes: 1. Processed dataset (`dataset_2channel_5s.npz`): 3,731 sliding windows from 388 unique records. - Shape: (N, 2, 10000) — channel 0 = PCG, channel 1 = ECG - Sampling rate: 2000 Hz; window length: 5 s; step: 3 s (2 s overlap) - Labels: 0 = Normal (1,072 windows), 1 = Abnormal (2,659 windows) - Preprocessing: ECG (50 Hz notch → 20 Hz low-pass → 0.4 Hz high-pass → z-score); PCG (25–400 Hz band-pass → z-score) - Includes `group_ids` and `source_records` for record-level cross-validation 2. Screening documentation: English criteria (`SCREENING_CRITERIA.md`), machine-readable build report (`build_report.json`), and discard list (`discarded_records.txt`). 3. Record-level discard gallery: 21 SVG waveform figures in three categories: - Insufficient channels (4 records) - No valid sliding-window segments (16 records) - NeuroKit2 ecg_peaks/ecg_delineate failure (1 record) Source data: Raw recordings are from the PhysioNet/CinC 2016 Challenge (training subset-a). Users must comply with the PhysioNet terms when accessing the original WFDB files. Build identifier: v20260623 Typical use: Multimodal heart-sound + ECG classification research, reproducible data screening, and record-grouped model evaluation. Loading example: ```python import numpy as np data = np.load('dataset_2channel_5s.npz', allow_pickle=True) X, y, groups = data['X'], data['y'], data['group_ids'] print(X.shape, y.shape) # (3731, 2, 10000) (3731,) Record-level 5-fold cross-validation can be performed with scikit-learn GroupKFold on the bundled group_ids field. A reference script (demo_5fold_group_split.py) is included; fixed fold indices are not provided.