ECG Cleaning¶

Unified cleaning interface with multiple backends.


`clean_ecg()`	Clean an ECG lead signal

ecgdatakit.processing.clean_ecg(lead, method='default', *, fs=None, **kwargs)[source]¶

Clean an ECG lead signal.

Parameters:

lead (Lead | ndarray[tuple[Any, ...], dtype[double]]) – Input ECG lead or raw signal array.
method (str) – Cleaning method: "default", "biosppy", "neurokit2", "combined", or "deepfade".
fs (int | None) – Sample rate in Hz. Required when lead is a numpy array.
**kwargs –
Extra arguments forwarded to the selected backend:
- device (str): PyTorch device for "deepfade" (default "cpu").
- weights_path (str | Path): Override the bundled DeepFADE weights.
- batch_size (int): Inference batch size for "deepfade" (default 32).

Returns:

Cleaned lead (new object, original unchanged).

Return type:

Lead

Available methods¶

Method	Extra dependency	Description
`"default"`	scipy	Bandpass 0.5–40 Hz + 50 Hz notch
`"biosppy"`	`pip install biosppy`	BioSPPy ECG filter
`"neurokit2"`	`pip install neurokit2`	NeuroKit2 adaptive pipeline
`"combined"`	biosppy + neurokit2	BioSPPy → NeuroKit2
`"deepfade"`	`pip install torch`	DeepFADE denoising autoencoder

DeepFADE¶

DeepFADE is a denoising autoencoder developed as part of ECGDataKit, trained on a large private multi-source ECG database with extensive noise augmentations (baseline wander, electrode motion, muscle artifacts, powerline interference). The architecture follows a symmetric DenseNet encoder-decoder design: the encoder compresses a 10-second single-lead ECG segment (500 Hz, 5 000 samples) through four dense blocks with progressive downsampling into an 8-channel latent representation, while the decoder mirrors the path with transposed-convolution upsampling and produces two outputs — the denoised signal and the estimated baseline wander. Pre-trained weights are bundled with the package.

from ecgdatakit.processing import clean_ecg

# CPU inference (default)
denoised = clean_ecg(lead, method="deepfade")

# GPU acceleration
denoised = clean_ecg(lead, method="deepfade", device="cuda")

# Apple Silicon MPS
denoised = clean_ecg(lead, method="deepfade", device="mps")

# Custom weights and batch size
denoised = clean_ecg(lead, method="deepfade", weights_path="my_weights.pt", batch_size=64)

Tip

Signals are automatically resampled to 500 Hz, segmented into 5 000-sample chunks, denoised in batches, and reassembled to the original length and sample rate.