PureWave

The PureWave project aims to develop an artifact detection algorithm in the synthetic electrocardiogram (ECG) data. Approximately 2,400 ECG data were generated using a CNN and Bi-LSTM model, based on actual ECG data from patients at Korea University Hospital (doi: 10.1007/s13755-023-00241-y).

Artifacts were classified visually according to clinical relevance and regarding the synthetic nature of the dataset, as follows:

• Wandering baseline

• High frequency artifact: AC interference & high frequency muscle tremor

• Medium frequency artifact: medium frequency & Parkinson’s disease muscle tremor

• Motion artifact

• Non-correlation of morphology: motion artifacts that only have PQRST morphology change

Lead-wise, time-dependent labels were manually generated at 0.2-second intervals.

An example of artifact labelling process.

We applied various data processing and feature extraction techniques, including band-pass filtering, wavelet transforms, and segmenting the data into shorter time intervals.

Multiple models were implemented, including deep learning-based approaches. However, likely due to the limited dataset size, conventional machine learning models with lower model complexity achieved the highest performance.

We are preparing our study results for presenting in the near conference.