ECG-Lense: Cardiac ECG Classification (IEEE Published)

Thu, 01 Aug 2024 00:00:00 +0000

Published at ETNCC 2024 · IEEE · DOI: 10.1109/ETNCC63262.2024.10767459

Headline numbers

80% classification accuracy
90% ROC-AUC, the best of all six models tested
21,837 ECG records · 18,885 patients · 12-lead · PTB-XL dataset

What it is

ECG-Lense is the deep CNN architecture I built and benchmarked against five other models for classifying cardiac ECG signals. Three classical ML, three deep learning, head-to-head on the same dataset and same metrics. ECG-Lense wins.

Model	Type
Decision Tree	Classical
Random Forest	Classical
Logistic Regression	Classical
Simple CNN	Deep
LSTM	Deep
ECG-Lense (Complex CNN)	Deep (proposed)

What I actually did

Worked off the PTB-XL dataset: clinical 12-lead ECGs, 10 seconds each, sampled at 100 Hz, with hierarchical diagnostic labels (5 superclasses, 24 subclasses)
Preprocessed signals with wfdb and standard ML libraries
Implemented a stationary wavelet transform (SWT) for data augmentation. The dataset is imbalanced and the augmentation increased sample diversity while preserving signal characteristics
Trained the deep models on raw ECG signals so feature extraction would be automatic
Evaluated on accuracy, precision, recall, F1, and ROC-AUC

What broke and what I learned

Class imbalance was the actual problem. First CNN runs hit a wall around 70% accuracy because rare cardiac classes were under-represented. SWT-based augmentation unlocked the next 10 points.
Raw signals beat hand-crafted features. Classical ML on extracted features stalled. The deep models, given raw signals, found patterns the feature engineering had thrown away.
More layers wasn’t the answer. ECG-Lense is “complex” relative to the simple CNN baseline, but each block earns its place. Depth for depth’s sake didn’t help.

Six months of work. A+ thesis grade. Submitted to IEEE during the project, accepted, presented and published at ETNCC 2024.

Stack

Python · TensorFlow · Keras · scikit-learn · MLFlow · wfdb · NumPy · pandas

Acknowledgements

Mentor: Dr. Kamal Kant Hiran (PhD). He pushed me to write the paper, guided the work end-to-end. Supervisors: Alok Kumar, Dr. Chandani Joshi, Harsh Bansal.

Dataset images

The dataset comprises 21,837 clinical 12-lead ECG records of 10 seconds in length from 18,885 patients. Hierarchical diagnostic organization: 5 superclasses, 24 sub-classes.
Metadata in ptbxl_database.csv: each ECG identified by a unique ecg_id with SCP codes for multi-label classification.

Demographic overview of patients in PTB-XL.

The dataset comprises 21837 clinical 12-lead ECG records of 10 seconds in length from 18885 patients. For the diagnostic labels, it provides a hierarchical organization in terms of 5 coarse superclasses and 24 sub-classes for the diagnostic labels
Columns provided in the metadata table ptbxl_database.csv. Each ECG is identified by a unique ID (ecg_id) and comes with several ECG statements (scp_codes) that can be used to train a multi-label classifier that can be evaluated based on the proposed fold assignments (strat_fold).

Demographic overview of patients in PTB-XL.

IEEE | Ukant Jadia