Deep learning optimisation for cardiology: Neural Architecture Search-driven arrhythmia classification with electrocardiograms
- 1Department of Engineering Science, The Podium Institute for Sports Medicine and Technology, University of Oxford, Oxford, United Kingdom
- 2Department of Engineering Science, Institute of Biomedical Engineering, University of Oxford, Oxford, United Kingdom
Abstract
Cardiovascular disease is the leading cause of death worldwide. Sudden cardiac death (SCD) accounts for roughly 50% of all cardiac deaths. The electrocardiogram (ECG) is widely used for diagnosing cardiac disease. However, the complexity of interpretation limits the ECG’s efficacy. Modern deep learning methods can assist clinicians in diagnosis. We applied Neural Architecture Search (NAS) to identify optimal deep learning architectures for classifying cardiac arrhythmias with ECGs. We applied the Differentiable Architecture Search strategy to an AutoFormer search and trained, validated, and tested the resulting model on the PhysioNet Challenge 2021 (PC21) dataset (n = 88,253 from three continents). We optimised hyperparameters on the NAS output, exploring input patch size, class weighting, and loss function. The PC21 metric and the area under the receiver operating characteristic curve (AUROC) are the performance metrics. The NAS converged towards minimal architectural configurations (embedding dimension: 384, depth: 4, self-attention heads: 4, MLP ratio: 1) with a validation and test PC21 metric of 0.66 (PC21 winner: 0.63). The NAS-created network achieved an AUROC of 0.97 and a PC21 metric of 0.72 during testing (PC21 winner: 0.69). We tested the PC21 winner architecture (1D-convolutional neural network with attention mechanism) without the attention mechanism and with the NAS-suggested architecture of 4 self-attention heads (instead of 8). The architecture without the attention mechanism performed similarly to the one with 8 self-attention heads (AUROC: 0.90, PC21 metric: 0.69). The architecture with 4 self-attention heads scored an AUROC of 0.91 and improved the PC21 metric to 0.71. The attention mechanism visualisation for two SCD-related arrhythmias (premature ventricular contractions and prolonged QT-interval) showed more clinically relevant attention for the 4-self-attention-head architectures. We interpret architectural simplicity as driving performance in arrhythmia classification. The self-attention mechanism can autonomously abstract clinical representations, simplifying clinical deployment by eliminating the need for an explicit feature-extraction pipeline.
Related
- Mechanistic Analysis and Optimisation of the Defluorination of Hydrofluorocarbon Refrigerants
- A geometry-matched structure generator outperforms basin hopping for silicon cluster global optimisation at fixed relaxation budget
- Critical assessment of published Kalina cycle studies: cycle configuration, state-point reporting and optimisation bounds
- Experimental Evaluation and Optimisation of Sealed‐Tube Combustion Methods for Silicate Nitrogen Extraction
- Development of an HPLC method for the determination of phenolic by-products: optimisation of the separation by means of the experimental designs methodology
BibTeX
@article{muller2026deep,
title = {{Deep learning optimisation for cardiology: Neural Architecture Search-driven arrhythmia classification with electrocardiograms}},
author = {Erik Vanegas Müller and Arese Joe-Oshodi and Abhirup Banerjee and Mauricio Villarroel},
journal = {PLOS One},
year = {2026},
month = nov,
doi = {10.1371/journal.pone.0362129},
url = {https://doi.org/10.1371/journal.pone.0362129},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Deep learning optimisation for cardiology: Neural Architecture Search-driven arrhythmia classification with electrocardiograms
AU - Erik Vanegas Müller
AU - Arese Joe-Oshodi
AU - Abhirup Banerjee
AU - Mauricio Villarroel
PY - 2026
DA - 2026/11/06
JO - PLOS One
DO - 10.1371/journal.pone.0362129
UR - https://doi.org/10.1371/journal.pone.0362129
AB - Cardiovascular disease is the leading cause of death worldwide. Sudden cardiac death (SCD) accounts for roughly 50% of all cardiac deaths. The electrocardiogram (ECG) is widely used for diagnosing cardiac disease. However, the complexity of interpretation limits the ECG’s efficacy. Modern deep learning methods can assist clinicians in diagnosis. We applied Neural Architecture Search (NAS) to identify optimal deep learning architectures for classifying cardiac arrhythmias with ECGs. We applied the Differentiable Architecture Search strategy to an AutoFormer search and trained, validated, and tested the resulting model on the PhysioNet Challenge 2021 (PC21) dataset (n = 88,253 from three continents). We optimised hyperparameters on the NAS output, exploring input patch size, class weighting, and loss function. The PC21 metric and the area under the receiver operating characteristic curve (AUROC) are the performance metrics. The NAS converged towards minimal architectural configurations (embedding dimension: 384, depth: 4, self-attention heads: 4, MLP ratio: 1) with a validation and test PC21 metric of 0.66 (PC21 winner: 0.63). The NAS-created network achieved an AUROC of 0.97 and a PC21 metric of 0.72 during testing (PC21 winner: 0.69). We tested the PC21 winner architecture (1D-convolutional neural network with attention mechanism) without the attention mechanism and with the NAS-suggested architecture of 4 self-attention heads (instead of 8). The architecture without the attention mechanism performed similarly to the one with 8 self-attention heads (AUROC: 0.90, PC21 metric: 0.69). The architecture with 4 self-attention heads scored an AUROC of 0.91 and improved the PC21 metric to 0.71. The attention mechanism visualisation for two SCD-related arrhythmias (premature ventricular contractions and prolonged QT-interval) showed more clinically relevant attention for the 4-self-attention-head architectures. We interpret architectural simplicity as driving performance in arrhythmia classification. The self-attention mechanism can autonomously abstract clinical representations, simplifying clinical deployment by eliminating the need for an explicit feature-extraction pipeline.
SN - 1932-6203
ER -