Deep-learning analysis of 12-lead ECGs detects drug-induced hERG inhibition and improves risk stratification beyond QTc
A Schubert, M Q Liang, L Sagers, Z ObermeyerAbstract
Background
Off-target inhibition of the human Ether-a-go-go-Related Gene (hERG) potassium channel can delay repolarisation and increase arrhythmia risk. Traditional monitoring based on heart rate-corrected QT interval (QTc) may miss cardiac risk signatures beyond interval prolongation.
Purpose
To develop and validate an artificial intelligence (AI) model that detects drug-induced hERG inhibition from routine 12-lead electrocardiograms (ECGs) and evaluate whether model predictions complement QTc-based risk stratification for arrhythmic outcomes.
Methods
A one-dimensional ResNet-18 was trained on ECG waveforms linked to medication administration from three US health systems (806,111 ECGs). Training used within-person pairs: baseline ECG within 30 days before drug initiation and on-treatment ECG within one drug-specific half-life after administration, with patient-level separation. Discrimination was assessed by area under the receiver operating characteristic curve (ROC-AUC; bootstrap 95% CI) for pre- vs post-drug ECGs in hERG and negative-control medications. External validation used ECGs from a randomised placebo-controlled crossover trial with paired plasma concentrations. Clinical relevance was assessed in chronic hERG-liability therapy users (>=2 consecutive medication orders; e.g. Sotalol, Ranolazine) by relating predicted risk to ventricular tachycardia and cardiac arrest, and by comparison with QTc.
Results
In the holdout set of paired pre-initiation and on-treatment ECGs, the model achieved ROC–AUC 0.847 (95% CI 0.844–0.850) for amiodarone/sotalol and 0.866 (0.857–0.876) for dofetilide, a direct hERG inhibitor withheld during training. Negative controls were near chance (0.576; 0.571–0.580). In external validation, predictions closely tracked plasma concentrations (Spearman ρ = 0.87–0.98; n = 15 timepoints per drug; all P ≤ 2.4×10−⁵), including dofetilide (ρ = 0.98), quinidine (ρ = 0.96), ranolazine (ρ = 0.87) and verapamil (ρ = 0.91). In patients receiving chronic hERG-liability therapies, predicted risk was associated with ventricular tachycardia and cardiac arrest. In linear probability models adjusting for QTc, each percentage point increase in predicted risk was associated with a 0.028 (95% CI: 0.026–0.030; P<0.001) percentage point increase in 30-day cardiac arrest rate. The 13.3% ECGs with highest predicted hERG liability showed 30-day cardiac arrest rates of 2.1% (vs. 0.7% in general population, p<0.001) with incomplete overlap with QTc>500 ms: 63.4% of high-risk ECGs were not flagged by QTc>500 ms. The highest 30-day cardiac arrest rate (2.65%) was observed in patients flagged by both approaches.
Conclusion
An AI model applied to serial 12-lead ECG waveforms detects drug-induced hERG inhibition, generalises to unseen drugs and trial data, and provides risk information complementary to QTc. This may enable scalable risk stratification and safety monitoring for QT-liable therapies in routine care.