Uncertainty estimation for deep learning-based automated analysis of 12-lead electrocardiograms
Problem
Deep neural networks for automated ECG interpretation show encouraging results, but whether such results can be trusted is paramount for clinical implementation.
Why it matters
Untrustworthy automated ECG readings could compromise clinical decisions and patient safety.
What the AI did
Systematically investigated six methods for estimating aleatoric and epistemic uncertainty in DNN-based ECG classification, evaluated on ranking, calibration, and out-of-distribution robustness, plus a clinical simulation with uncertainty thresholds.
Results
526,656 ECGs from three datasets; six uncertainty methods evaluated on ranking, calibration, and out-of-distribution robustness; clinical simulation applying increasing uncertainty thresholds.
Impact
Provides systematic evidence on uncertainty estimation techniques for trustworthy automated ECG analysis.
Record
| Venue | European Heart Journal - Digital Health |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Jeroen F Vranken, Rutger R van de Leur, Deepak Kumar Gupta, Luis Eduardo Juárez‐Orozco, Rutger J. Hassink, Pim van der Harst, Pieter A. Doevendans, Sadaf Gulshad et al. |
| Citations | 41 (OpenAlex, at import time) |
| References | 16 |
| Model or method | deep neural networks (DNNs) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1093/ehjdh/ztab045 |
| Open access | gold |
| Keywords | electrocardiogram, deep learning, uncertainty estimation, calibration, out-of-distribution data |
Importance rating
significant — thousands to millions affected
Trustworthiness is paramount for clinical implementation of automated ECG reading; large-scale evaluation across three datasets supports patient-safety relevance.