Uncertainty-Quantified Gaussian Process Classification of Torsades de Pointes Risk from Pseudo-Coloured ECGs: External Validation in a UK Biobank Breast Cancer Cohort

Abstract Machine-learning models for electrocardiogram (ECG)-based risk assessment are usually evaluated only on their development population, so their behaviour under population shift is rarely known. We examined this problem for Torsades de Pointes (TdP), a life-threatening arrhythmia associated with delayed ventricular repolarisation. A Bayesian Gaussian Process Classifier (GPC) with an Automatic Relevance Determination (ARD) Matérn (ν = 2.5) kernel was trained to classify TdP risk, with labels defined by the QT nomogram, from 25 pseudo-coloured ECG features of healthy volunteers receiving QT-prolonging drugs (ECGRDVQ; 5,219 ECGs, 180 at risk of TdP) and applied without retraining to an independent UK Biobank cohort of patients with breast cancer (2,478 ECGs, 59 at risk of TdP). This population was chosen because it is commonly exposed to cardiotoxic, QT-prolonging therapies and is predominantly female, an established TdP risk factor. We also applied explainability analyses and decomposed predictive uncertainty. Internally, Balanced Accuracy was 0.987, sensitivity 0.978 and specificity 0.997. Externally, Balanced Accuracy was 0.835, sensitivity 0.814 and specificity 0.856. ARD length-scales, permutation importance and SHapley Additive exPlanations (SHAP) identified a single dominant feature, InflectionPoint, whose specificity declined externally: ECGs without TdP risk with InflectionPoint ≥ 4.5 rose from 7.3% to 15.1%. Aleatoric uncertainty exceeded epistemic uncertainty (0.167 versus 0.115 nats), indicating that errors arose mainly from class overlap rather than limited model knowledge. The model retained most of its sensitivity externally, and the analyses showed why specificity fell and what deployment requires: more discriminative features and population-representative training data.