Searching for Premature Ventricular Contraction and Supraventricular Premature Beat from Long-term ECGs: The 3rd China Physiological Signal Challenge 2020
Tests are done on a 60s segment (median-filtered and bandpassed, sample1_fs250.mat in this folder) of a subject with frequent PVC.
- r peak detections are done using this function.
- PVC beats are labeled using red vertical lines.
- missed PVC beats are labeled using yellow boxes.
- the first image is the result by a modified version of machine learning algorithms from this repo using rr features and wavelet features, with post-processing using clinical rules. Note that phase_one_legacy is one such modified version, which uses XGBoost instead of SVM and without clinical post-processing.
- the second image is the result of the sequence labeling deep learning model with probability threshold 0.3, and filtered by a deep learning classifier. The missed PVC beats are caused by this classifier.
- the last image is the result of the sequence labeling deep learning model with probability threshold 0.5, and filtered by a deep learning classifier.
- an effective and robust rpeak (qrs complex) detector is crucial.
- the sequence labeling deep learning model (trained only for a dozen epochs because of the approaching deadline) tends to make false positive predictions but seldom has false negatives; while the deep learning classifier (trained only for several hundred epochs) has few false positives but has slightly higher probability to have false negatives.
- given a good rpeak detector, machine learning models might well be competitive against deep learning models.
- changing the threshold of the sequence labeling deep learning model from 0.5 to 0.3 can largely reduce the PVCerr score (punishment); further removing the post-filtering of the deep learning classifier might further reduce the scores, raising more false positives while reducing false negatives, considering that false negative has punishment 5 times as the punishment of false positives.
Evaluation result on the final full hidden test set
not fully listed
[2] BioSPPy
[3] Cai, Wenjie, and Danqin Hu. "QRS complex detection using novel deep learning neural networks." IEEE Access (2020).
[4] torch_ecg
[1] more robust qrs detector (finished)
[2] feature engineering (deprecated)
[3] deep learning model structure design (ongoing)
[4] use SNR to deal with (eliminate?) too noisy segments?
[5] etc....
if you find this function useful, please cite Reference [3]