[Paper Review] Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks

Pranav Rajpurkar, Awni Hannun+3 more|arXiv (Cornell University)|Jul 6, 2017

ECG Monitoring and AnalysisMedicine21 references618 citations

TL;DR

A 34-layer CNN trained on a 30,000-patient, 14-class single-lead ECG dataset achieves cardiologist-level performance and often surpasses average cardiologists in detecting a range of arrhythmias.

ABSTRACT

We develop an algorithm which exceeds the performance of board certified cardiologists in detecting a wide range of heart arrhythmias from electrocardiograms recorded with a single-lead wearable monitor. We build a dataset with more than 500 times the number of unique patients than previously studied corpora. On this dataset, we train a 34-layer convolutional neural network which maps a sequence of ECG samples to a sequence of rhythm classes. Committees of board-certified cardiologists annotate a gold standard test set on which we compare the performance of our model to that of 6 other individual cardiologists. We exceed the average cardiologist performance in both recall (sensitivity) and precision (positive predictive value).

Motivation & Objective

• Motivate automatic, accurate arrhythmia detection from single-lead ECGs to aid clinical workflows.
• Create a dataset orders of magnitude larger than prior corpora to enable deep learning.
• Develop a fully end-to-end CNN model that maps raw ECG time-series to rhythm annotations.
• Compare model performance to board-certified cardiologists using gold-standard annotations.

Proposed method

• Use a 34-layer convolutional neural network with residual connections to map 30-second, 200 Hz ECG segments to per-second rhythm labels.
• Train end-to-end with robust normalization and dropout; employ Adam optimizer with learning-rate scheduling.
• Annotate a large ECG dataset (64,121 records from 29,163 patients) with 14 rhythm classes by clinical experts; test against a committee of cardiologists for ground truth.
• Output predictions at 1 Hz from 30-second inputs, enabling sequence-to-sequence labeling of rhythms.

Experimental results

Research questions

• RQ1 Can a deep CNN outperform average board-certified cardiologists in detecting a broad set of arrhythmias from single-lead ECGs?
• RQ2 Does increasing dataset size and model depth improve sequence-to-sequence rhythm labeling accuracy on ECGs?
• RQ3 How does model performance compare to clinicians on both per-class F1 and aggregate metrics (precision, recall, F1) across 14 rhythm classes?

Key findings

• The model outperforms the average cardiologist on both sequence-level F1 and set-level F1 metrics.
• Per-class F1 scores show the model achieving higher F1 than cardiologists on many rhythms, notably AV block types and CHB.
• Aggregate precision, recall, and F1 favor the model over cardiologists in both sequence and set evaluations.
• A large dataset (≈64k records from ≈29k patients) and a 33-layer CNN with residual connections are key for performance gains.
• The model often confuses similar rhythms in line with clinical similarities (e.g., Wenckebach vs Mobitz II), illustrating reasonable error patterns.