Abstract
Background – Many medications are associated with long QTc (LQT). Current LQT predictors have limited generalizability and/or modest performance.

Objective – To compare performance of machine learning-enabled approaches to drug-induced LQT prediction with current risk scores, including Tisdale and RISQ-PATH.

Methods – We identified patients who began QTc-prolongers (QTdrugs) and had follow-up 12-lead ECGs within 1 year. Using 5-fold cross-validation, we trained XGBoost (XGB), ECG-based deep neural network (DNN), and combined models using EHR data and ECG traces to predict QTc ≥500 ms within 1 year. We assessed performance using areas under the receiver operating characteristic (AUROC) curves and compared against corresponding Tisdale and RISQ-PATH scores.

Results – QTdrug records were identified for 345,371 patients, 5.7% of whom had LQT within 1 year. In the subset with baseline ECGs available (N=182,448; 7.7% events), both the XGB and DNN models demonstrated high performance (AUROC=0.869 and 0.864, respectively) but their combination yielded no significant improvement (AUROC=0.874). Therefore, focusing on the XGB model, we observed superior performance vs. RISQ-PATH in the overall population (AUROC=0.859 vs. 0.701), as well as Tisdale in predominantly inpatients (N=110,558; 8.8% events; AUROC=0.855 vs. 0.770). Positive predictive value was 61.5% vs. 28.35% and 54.0% vs. 28.3% at equivalent operating points for the XGB model vs. RISQ-PATH and Tisdale, respectively.

Conclusion – Development and retrospective validation of three machine learning-based models for predicting drug-induced LQT at the time of new QTdrug starts demonstrated superior performance compared with current clinical risk calculators and may be useful tools to support medical decision-making when initiating new therapies.

VIEW THE PUBLICATION