Background: Drug-induced QTc prolongation (diQTP) is frequent and associated with a risk of sudden cardiac death . Identifying patients at risk of diQTP can enhance monitoring and treatment plans .

Objective: To develop a machine learning architecture for prediction of extreme diQTP (QTc> 500ms OR {Delta} QTc> 60ms) at the onset of treatment with a QTc prolonging drug .

Methods: We included 4,628 adult patients who received a baseline ECG within 6 months prior to treatment onset with a QTc prolonging drug and a follow-up ECG after the fifth dose . We collected known clinical QTc prolongation risk factors (CF). We developed a novel neural network architecture (QTNet) to predict diQTP from both the CF and baseline ECG data (ECGD), composed of both the ECG waveform and measurements (i.e . QTc), by fusing a state-of-the-art convolution neural network to process raw ECG waveforms with the CF using amulti-layer perceptron . We fit a logistic regression model using the CF, replicating RISQ-PATH as Baseline . We further compared the performance of QTNet (CF+ECGD) to neural network models trained using three variable subsets: a) baseline QTc (QTC-NN), b) CF-NN, and c) ECGD-NN .

Results: diQTP was present in 1030 patients (22.3 %), of which baseline QTc was normal (QTc <450ms: Male/ <470ms: Female) in 405 patients (39.3 %). QTNet achieved the best performance (Figure 1) (AUROC , 0.802 [95% CI , 0.782-0.820] ), outperforming predictions based on the Baseline (AUROC , 0.738 [95% CI , 0.716-0.757] ), QTC-NN (AUROC , 0.735 [95% CI , 0.710-0.757] ), CF-NN (AUROC , 0.778 [95% CI , 0.757-0.799] ), and ECGD-NN (AUROC , 0.774 [95% CI , 0.750-0.794] ). Conclusion: We developed QTNet, the first deep learning model for predicting extreme diQTP, outperforming models trained on known clinical risk factors.