What is the clinical evidence from this study?

Study design: Other. Population: Arrhythmia / Electrocardiogram feature extraction (n=12989). Intervention: Convolutional variational autoencoder (CVAE) vs. Random initialization. Primary outcome: f1-score for arrhythmia classification.

December 1, 2021PLoS ONE39 citationsOpen Access

Unsupervised feature learning for electrocardiogram data using the convolutional variational autoencoder

JJJong-Hwan JangSeoul Medical Center TKTae Young KimInje University HLHong‐Seok LimAjou University

Key Result

Unsupervised feature learning with a convolutional variational autoencoder extracted meaningful ECG characteristics, achieving an f1-score of 0.86 for arrhythmia classification using CVAE features alone.

Structured PICO

Does unsupervised feature learning using a convolutional variational autoencoder (CVAE) improve ECG feature extraction and arrhythmia classification compared to random initialization?

Population

12,989 patients from multiple databases (ICU, medical check-up, and public datasets) whose ECG data were used to train and validate an unsupervised convolutional variational autoencoder for feature extraction.

Intervention

Unsupervised feature learning using a convolutional variational autoencoder (CVAE) trained on ECG lead II data.

Comparator

Random initialization (for transfer learning models).

Outcome

Performance of CVAE features in clustering, anomaly detection, and transfer learning for arrhythmia classification (measured by f1-score, accuracy, precision, and recall).surrogate

Unsupervised feature learning with a convolutional variational autoencoder can effectively extract characteristics of various ECG rhythms and improve arrhythmia classification performance via transfer learning.

Limitations

The CVAE was developed using data from a single cohort of patients from the AUMC ICU.
The length of the ECG was static at 8.2 seconds because it depended on the model structure.
Only ECG lead II was trained because most ICUs mainly monitored and collected ECG lead II.
The proposed threshold for anomaly detection may not be applicable in all circumstances.
Each feature value does not correspond to a specific clinical change, requiring further investigation for interpretation.

Abstract

Most existing electrocardiogram (ECG) feature extraction methods rely on rule-based approaches. It is difficult to manually define all ECG features. We propose an unsupervised feature learning method using a convolutional variational autoencoder (CVAE) that can extract ECG features with unlabeled data. We used 596,000 ECG samples from 1,278 patients archived in biosignal databases from intensive care units to train the CVAE. Three external datasets were used for feature validation using two approaches. First, we explored the features without an additional training process. Clustering, latent space exploration, and anomaly detection were conducted. We confirmed that CVAE features reflected the various types of ECG rhythms. Second, we applied CVAE features to new tasks as input data and CVAE weights to weight initialization for different models for transfer learning for the classification of 12 types of arrhythmias. The f1-score for arrhythmia classification with extreme gradient boosting was 0.86 using CVAE features only. The f1-score of the model in which weights were initialized with the CVAE encoder was 5% better than that obtained with random initialization. Unsupervised feature learning with CVAE can extract the characteristics of various types of ECGs and can be an alternative to the feature extraction method for ECGs.

Ask AI

Helpful

Bookmark

View Full Paper