Deep Learning-Based Continuous QT Monitoring to Identify High-Risk Prolongation Events After Class III Antiarrhythmic Initiation

We thank Li and colleagues (this issue) for their thoughtful comments, which appropriately distinguish between technical feasibility, clinical interpretation, and implementation.¹ Our study² was intended to establish that a single nonstandard ambulatory vector can support clinically meaningful QT monitoring while identifying a clear path for prospective deployment.

Regarding validation of 3DRECON-QT, we agree that larger prospective studies using true insertable cardiac monitor (ICM) recordings are needed. The true-ICM cohort was limited by the scarcity of paired proprietary ICM waveforms and simultaneous 12-lead ECGs in this non-industry-sponsored work. Accordingly, the V3-V2 surrogate strategy served as a translational bridge where large-scale simultaneous training data do not exist. Importantly, support for generalizability did not rely on the true-ICM cohort alone, but also on concordant findings across held-out internal, external, continuous-monitoring, and high-risk subset analyses. The close agreement between derived and true ICM predictions supports this training approach.

We wish to clarify that the study did not use “≥500 ms or ≥15% increase from baseline” as a single combined definition of high-risk prolongation. The primary endpoint across cohorts was the established absolute threshold derived from dofetilide labeling, whereas the ≥15% increase was examined only as a secondary within-patient analysis in the continuous-monitoring cohort. These analyses were reported separately to reflect distinct clinical questions: absolute risk threshold and dynamic temporal change. Current thresholds for identifying high-risk QT prolongation derive from inpatient loading protocols, and the appropriate threshold for long-term, ambulatory QT monitoring in patients on Class III antiarrhythmics remains an open question in the absence of technology enabling its study.³

We acknowledge that the encounter-based outpatient cohort enriches for clinically complex patients relative to all ambulatory users of class III antiarrhythmics. The observed prevalence should not be interpreted as a population-wide estimate, but as broad and clinically relevant high-risk cohort including all cardiac and noncardiac-related acute care encounters. We also respectfully clarify that the multivariable analysis did account for several confounders raised by Li and colleagues, including creatinine, potassium, and magnesium, in addition to age, heart failure, atrial fibrillation, and drug class. Concerning ventricular arrhythmias, the low absolute event rate warrants cautious interpretation of effect size, although rates were comparable to prior studies.⁴ Our intent was not to define a precise causal estimate, but to show that prolongation identified by the model is clinically meaningful rather than a measurement artifact.

Finally, our use of the term “continuous QT monitoring” represents practical requirements for medication safety rather than edge deployment of the model. Contemporary ICM platforms acquire and transmit up to hourly intervals, which is well matched to the pharmacokinetic timescale of Class III antiarrhythmic-induced QT change.⁵ Although not demonstrated in this study, cloud-based deployment of 3DRECON-QT between hourly collection on device and the physician dashboard is readily feasible and could provide meaningfully enhanced monitoring for the affected patients. In our view, the broader significance of this work is that it establishes a scalable framework for spatial reconstruction from sparse biosignals, particularly in settings where simultaneous ground-truth training data are limited.