Source-linked AI summary
Arrhythmia Classification from the Abductive Interpretation of Short Single-Lead ECG Records
Tomás Teijeiro, Constantino A. García, Daniel Castro, Paulo Félix
TL;DR
Automatic detection of atrial fibrillation from short single-lead ECGs remains a challenge, while the potential of low-cost ECG tests for early cardiac-disease detection is still largely untapped. The paper combines knowledge-based interpretation with learning-based classification using high-level, clinically meaningful ECG features. It reports that XGBoost improves on interpretation alone and that combining knowledge-based and learning-based approaches can retain interpretability.
Problem
The potential of low-cost ECG tests for early cardiac-disease detection remains largely untapped, and the 2017 challenge seeks automatic atrial-fibrillation detection from short single-lead ECGs.
Source: §1 · p. 1Method
The approach uses abductive interpretation and a domain-specific knowledge base to describe ECGs with high-level features used by cardiologists, then combines knowledge-based and learning-based classification.
Source: §1 · p. 1, §2.2 · p. 2, §4 · p. 4Results
The authors report that XGBoost improves results over interpretation alone and that combining knowledge-based and learning-based approaches is effective while maintaining interpretability through high-level meaningful features.
Source: §2.4 · p. 3, §4 · p. 4Takeaways & Limitations
Formalizing anomalies such as tachycardia, bradycardia, and wide QRS complexes helped optimize the training set by defining more specific features from interpretation results.
Source: §2.4 · p. 3Some training labels appear inconsistent: records with PR intervals over 210 ms occur in the normal class, while atrial-flutter patterns occur in records labeled atrial fibrillation.
Source: §2.4 · p. 3
Abstract
from arXiv · showhide
In this work we propose a new method for the rhythm classification of short single-lead ECG records, using a set of high-level and clinically meaningful features provided by the abductive interpretation of the records. These features include morphological and rhythm-related features that are used to build two classifiers: one that evaluates the record globally, using aggregated values for each feature; and another one that evaluates the record as a sequence, using a Recurrent Neural Network fed with the individual features for each detected heartbeat. The two classifiers are finally combined using the stacking technique, providing an answer by means of four target classes: Normal sinus rhythm, Atrial fibrillation, Other anomaly, and Noisy. The approach has been validated against the 2017 Physionet/CinC Challenge dataset, obtaining a final score of 0.83 and ranking first in the competition.
1. Introduction
The paper addresses automatic atrial-fibrillation detection in short, single-lead ECGs, while emphasizing interpretability as important for clinical trust. It describes a knowledge-based, cardiologist-like signal representation that machine-learning classifiers use to address limitations in the training labels and interpretation.
1. Introduction
- The 2017 challenge frames atrial-fibrillation detection from short single-lead ECGs as a supervised classification task evaluated on a hidden test set.The authors also note interpretability as a concern for clinical adoption and staff trust.Source: §1 · p. 1
- The proposal represents ECG signals with high-level features also used by cardiologists in ECG analysis.The representation is generated using a knowledge-based approach.Source: §1 · p. 1
- An abductive framework selects explanatory hypotheses for observed evidence before machine learning addresses incomplete labeling criteria and possible interpretation errors.The framework is for time-series interpretation.Source: §1 · p. 1
2. Methods
The method combines abductive interpretation of ECG evidence with global and per-beat features, then uses classifiers and stacking to assign rhythm labels. Preprocessing addresses label inconsistencies and inverted records, while the interpretation can revise initial beat evidence.
2.1. Preprocessing
- Preprocessing manually relabeled 197 of 8,528 training records and used a logistic-regression classifier to detect inverted records after manual identification.The inversion classifier used 14 features based on raw signals and tentative waveform delineation.Source: §2.1 · p. 1
2.2. Abductive interpretation
- The interpretation stage uses Construe to build rhythm hypotheses from waveform evidence and can discard unsupported beats or seek missed beats predicted by a selected pattern.This non-monotonic correction can remove false-positive beats from noisy or artifact-affected signals.Source: §2.2 · p. 2
2.3. Global feature extraction
- The 79 global features summarize rhythm, morphology, and signal quality, including measures of intervals, waveform properties, and profiled signal regions.Signal-quality features assess the importance of morphological features indicating conduction anomalies.Source: §2.3 · p. 2
2.4. Global classification
- The XGBoost classifier learns class criteria and possible interpretation errors because challenge labeling guidelines, especially for the Other anomaly class, are unavailable.Its hyperparameters were selected by grid search and 8-fold cross-validation.Source: §2.4 · p. 2
- The authors formalized anomalies for classifying records as O, including tachycardia (mean heart rate over 100 bpm), bradycardia (mean heart rate under 50 bpm), wide QRS complexes (longer than 110 milliseconds), ventricular or fusion beats, and at least one extrasystole.Source: §2.4 · p. 3
2.5. Per-beat feature extraction
- Global features can miss episodic abnormalities, so the sequence classifier uses per-beat morphology, intervals, RR variation, and P-wave profiles to learn temporal patterns.The motivating example is a normal record containing a single ectopic ventricular beat that does not break the rhythm.Source: §2.5 · p. 3
2.6. Sequence classification
- The sequence model uses LSTMs to process transformed per-beat features and outputs class probabilities for four classes.The architecture also uses temporal mean and maximum pooling; its outputs are combined through an MLP and Softmax layer.Source: §2.6 · p. 3, Figure 3 · p. 4
2.7. Classification stacking
- The stacking stage combines XGBoost predictions with predictions averaged across three RNNs using a Linear Discriminant Analysis classifier.Only three probabilities from each model are used to avoid possible collinearity issues.Source: §2.7 · p. 4
3. Evaluation
Evaluation follows the challenge guidelines and uses the mean F1 measure of the N, A, and O classes as the final score. An example table reports stratified 8-fold cross-validation results.
3. Evaluation
- The final score is the mean F1 measure of the N, A, and O classes, and Table 1 gives an example of stratified 8-fold cross-validation results.The authors state that the stacker usually scores better than the base models and has lower variance, though the variance is not shown in the table.Source: §3 · p. 4, Table 1 · p. 4
4. Conclusions
The work presents knowledge-based and learning-based methods as an effective combination for classification, with interpretability retained through high-level, meaningful features.