Source-linked AI summary
Weak Supervision for Affordable Modeling of Electrocardiogram Data
Mononito Goswami, Benedikt Boecking, Artur Dubrawski
TL;DR
Manual annotation makes training ECG arrhythmia models costly despite the availability of unlabeled data. The paper uses domain-designed time-series heuristics to create probabilistic labels without pointwise ground truth, and reports competitive models using only a handful of heuristics. The illustrated work focuses on PVCs, though the authors state the approach applies to all abnormal-heartbeat classes.
Problem
ECG arrhythmia models rely on large manually annotated datasets, but point-by-point heartbeat labeling is tedious and expensive.
Source: Introduction · p. 1Method
The authors use domain-designed, potentially noisy heuristics as labeling functions and model them to produce probabilistic heartbeat labels without individual ground-truth labels.
Source: Introduction · p. 1, Methodology · p. 4Results
With six heuristics and little human effort, the model was competitive with full ground-truth training and outperformed a model trained using active learning.
Source: Introduction · pp. 1–2, Experiments and Results · p. 6Takeaways & Limitations
The authors conclude that direct time-series heuristics offer a promising way to train medical ML models without large manually annotated datasets.
Source: Discussion and Conclusion · p. 8The experiments illustrate PVC detection, although the authors state the approach applies to all abnormal-heartbeat classes.
Source: Introduction · p. 2
Abstract
from arXiv · showhide
Analysing electrocardiograms (ECGs) is an inexpensive and non-invasive, yet powerful way to diagnose heart disease. ECG studies using Machine Learning to automatically detect abnormal heartbeats so far depend on large, manually annotated datasets. While collecting vast amounts of unlabeled data can be straightforward, the point-by-point annotation of abnormal heartbeats is tedious and expensive. We explore the use of multiple weak supervision sources to learn diagnostic models of abnormal heartbeats via human designed heuristics, without using ground truth labels on individual data points. Our work is among the first to define weak supervision sources directly on time series data. Results show that with as few as six intuitive time series heuristics, we are able to infer high quality probabilistic label estimates for over 100,000 heartbeats with little human effort, and use the estimated labels to train competitive classifiers evaluated on held out test data.
Introduction
ECG arrhythmia detection can reduce reliance on difficult manual review, but supervised models depend on costly point-by-point annotation. This work tests whether noisy domain heuristics can provide probabilistic labels for competitive arrhythmia models.
Introduction
- Manual ECG review is cumbersome and error-prone, while large supervised models require laborious annotation of extensive datasets.The passage notes that reliance on labeled data can leave researchers using static, older datasets despite changing patients, disease understanding, and equipment.Source: Introduction · p. 1
- The approach uses cheaper, potentially noisier supervision sources to train an arrhythmia detector without ground-truth labels for individual samples.Experts encode domain knowledge as noisy labeling functions; a label model estimates their accuracy and produces probabilistic labels through weighted majority vote.Source: Introduction · p. 1, Figure 1 · p. 2
- With six heuristics and little human effort, the model was competitive with full pointwise ground-truth training and outperformed an active-learning model.The experiments used ECG data from the MIT-BIH Arrhythmia Database.Source: Introduction · pp. 1–2
- The paper illustrates the method with Premature Ventricular Contractions, while stating that it applies to all classes of abnormal heartbeats.Frequent PVCs with exceptionally wide QRS complexes may indicate heart disease and eventually lead to sudden cardiac death.Source: Introduction · p. 2
Related Work
Prior arrhythmia models commonly rely on manually labeled ECG corpora, while semi-supervised and active-learning approaches seek to reduce expert labeling. This work places its contribution within weak supervision by defining heuristics directly on time series.
Automated Arrhythmia Detection
- Automated arrhythmia detection has often used manually labeled corpora, and a cited CNN trained on 64,121 records from 29,163 patients used technician annotations.The section frames affordable labeling of large data volumes as a need for progress in automated arrhythmia detection and ML-aided healthcare.Source: Related Work · p. 2
- Semi-supervised and active-learning studies reduce expert intervention in different settings; Pasolli et al.14 used three strategies to select heartbeats for an SVM.This paper also compares its weakly supervised method with active learning, using a CNN with uncertainty sampling rather than curated ECG features and margin sampling.Source: Related Work · pp. 2–3
Weak Supervision
- Unlike most prior data-programming work on images or natural language, this paper defines heuristics directly on time series for labeling whole series or segments.Prior work often used technician or student annotations, auxiliary modalities, or labels only at the whole-series level.Source: Related Work · p. 3
Methodology
The method turns six time-series heuristics into probabilistic heartbeat labels, then trains a noise-aware classifier on those estimates. It personalizes heuristic thresholds by patient and evaluates the classifier on held-out data.
Methodology
- Six domain heuristics label beat subsets; a factor-graph label model estimates probabilistic labels that train a noise-aware ResNet classifier.The classifier weights each sample by its maximum probability of belonging to either class.Source: Methodology · pp. 3–5
- The six PVC heuristics use ECG morphology and timing, including early, tall, or wide R-waves and QRS/ST-vector direction.A non-domain expert defined the time-series heuristics in less than 30 minutes.Source: Methodology · p. 3
- The factor graph models LF accuracy and propensity and assumes labeling functions are conditionally independent given the unobserved class label.Snorkel learns the model parameters from observed LF outputs, producing posterior labels for classifier training.Source: Methodology · pp. 3–4
- ECGs are minimally filtered, segmented into beats between alternate QRS-complexes, and analyzed to locate R/T waves and estimate the baseline.RANSAC fits a robust linear regression baseline, while peak finding locates the R and T waves.Source: Methodology · p. 4, Figure 2 · p. 4
- Patient-specific thresholds turn qualitative rules such as an unusually early R-wave into labeling functions, with some heuristics requiring no subject-specific parameters.Thresholds are computed separately for each patient using a robust estimate and a two-standard-deviation cutoff in the direction of interest.Source: Methodology · p. 4, Figure 3 · p. 5
Experiments and Results
On held-out MIT-BIH test data, the weakly supervised ResNet performed on par with the fully supervised model and outperformed the active-learning model. Results also highlight the value of subject-specific heuristic tuning and show that the two ResNets focus on similar QRS-complex features.
Experimental Setup
- The evaluation used the first channel of 48 half-hour, two-lead recordings from 47 subjects, with an inter-patient split and PVC oversampling in training.The ResNet models were trained for 25 epochs; the AL baseline incrementally labeled 4,000 points from a balanced 100-point seed set.Source: Experiments and Results · pp. 5–6
Results
- The weakly supervised ResNet performed on par with the fully supervised ResNet and outperformed the ResNet trained using active learning on held-out test data.Models were trained on DS1 and evaluated on DS2.Source: Table 1 · p. 6, Experiments and Results · p. 6
- At up to 4,000 labeled points, the weakly supervised model matched or exceeded its active-learning counterpart with substantially less human effort.Figure 5 reports the active-learning comparison; its shaded red regions show 95% Wilson score intervals.Source: Figure 5 · p. 6, Experiments and Results · p. 7
- Automatically tuned labeling-function thresholds varied substantially across subjects; the R-wave position threshold had a mean of 230 ms and standard deviation of 77.14 ms.The model performed poorly when these parameters were fixed to global settings; probabilistic-label sensitivity also varied from 0.994 for Subject 228 to 0 for Subject 214.Source: Experiments and Results · pp. 6–7
- Class Activation Maps showed both weakly and fully supervised ResNets focusing primarily on QRS-complex morphology in the illustrated PVC and other beats.The models appeared to focus on similar signal regularities.Source: Figure 6 · p. 7, Experiments and Results · p. 7
- At low-error operating settings, the weakly supervised ResNet was particularly strong at identifying negatives, while positive recall was close to the fully supervised equivalent.The inferred probabilistic labels alone had little operational utility.Source: Experiments and Results · p. 7
Discussion and Conclusion
Weak supervision with time-series domain heuristics produced an arrhythmia model on par with its fully supervised counterpart, without point-by-point annotation. The authors also identify patient-specific tuning and broader clinical use as implications and directions for future work.
Discussion and Conclusion
- The weakly supervised arrhythmia model performed on par with its fully supervised counterpart without point-by-point annotation and was developed in a fraction of the time.The authors describe time-series domain heuristics as a promising route to training medical ML models without large manually annotated datasets.Source: Discussion and Conclusion · p. 8
- A handful of intuitive heuristics yielded reliable probabilistic labels, and parameter-tuning rules accounted for inter-patient variability while keeping manual labor minimal.The authors expect additional heuristics harvested from expert clinicians could further improve results.Source: Discussion and Conclusion · p. 8
- Existing alternatives include active learning, semi-supervised learning, and crowdsourcing, but semi-supervised arrhythmia methods still require labeled data and crowdsourcing has reported limitations.Reported crowdsourcing concerns include trustworthiness, fine-grained annotation by non-experts, and patient privacy.Source: Discussion and Conclusion · p. 8
- Future work includes modeling dependencies between labeling functions, building time-series primitives and interactive tools, validating end models, and testing hybrid, clinical, and broader waveform applications.The authors also propose studying other ECG abnormalities and hemodynamic monitoring waveforms.Source: Discussion and Conclusion · p. 8
- The authors argue that reducing data-preparation and maintenance costs could support wider adoption of machine learning for clinical decision support using healthcare time series.They frame this as a potential contribution of approaches like the one presented, not a demonstrated outcome.Source: Discussion and Conclusion · p. 8