Bringing doctors together

or

A free, private network for verified physicians

Leading voices

Dr. Pugazhendhi Vijayaraman

EP · Geisinger

View

Dr. Robert Schaller

EP · HUP

View

Dr. Felix Yang

EP · NYU Langone

View

Dr. Gregory Marcus

EP · UCSF

View

Dr. Miguel Leal

EP · Emory University Hospital

View

Dr. Chirag Barbhaiya

EP · NYU

View

Dr. William Sauer

EP · BWH

View

Dr. Dorairaj Prabhakaran

General Cardiology

View

Dr. Trey Woods

General Cardiology · Emory Healthcare

View

Dr. Brahmajee Nallamothu

Interventional Cardiology · UMHS

View
  • Discover your community

  • Stay current

  • A profile that represents you

  • Find your people

About RoonPrivacyTerms

    Overview

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

    Nov 8, 2025DOI: 10.1161/circulationaha.125.077494Journal ArticleAHA

    Written by: Rayan A AnsariSabyasachi BandyopadhyayRishi K TrivediKelly A BrennanXichong LiuPrasanth GanesanJ Weston HughesAlexander C PerinoEuan A AshleyPaul J WangTodd ColemanMarco V PerezDavid OuyangSanjiv M NarayanAlbert J Rogers

    Find these authors on Roon

    PW
    Dr. Paul Wang
    EP · Stanford

    METHODS

    Model development was performed at Stanford University and tested at an independent external site (Cedars-Sinai Medical Center). The study was approved by the institutional review boards of both centers. A waiver of consent was granted for retrospective analysis. Across these sites and a public database, 5 distinct patient cohorts were established. These comprised the Stanford Model Development Cohort, Cedars-Sinai External Validation Cohort, Continuous QT Monitoring Cohort (ECG Effects of Ranolazine, Dofetilide, Verapamil, and Quinidine PhysioNet Dataset), Stanford Real-World Class III Antiarrhythmic Exposure Cohort, and the Stanford ICM Recording Cohort. The data that support the findings of this study are available from the corresponding author upon reasonable request. A dataset overview and model input for each analysis is available in Table S1.We analyzed 794 813 12-lead ECG data obtained from 258 205 unique patients in the Stanford Hospital ECG database (iECG, Phillips, Amsterdam, The Netherlands). The data collection period extended from 2005 to 2018. Patients were randomly divided into training (80%), validation (10%), and testing (10%) cohorts to ensure all ECGs from any single patient were assigned exclusively to one data partition. The test set (10%) included 78 245 12-lead ECG recordings from 25 821 patients not previously seen by the model during training. Exclusion criteria included records in legacy formats before system migration, poor signal quality (>70% of leads 0-valued or flatlined), incomplete or absent machine annotations, and missing QT or RR interval measurements. All rhythms, including atrial and ventricular paced ECGs, were included. The consort diagram for the development and validation cohorts is provided in Figure 1.ECGs (N=1 048 575) from a geographically distinct health system and alternate ECG acquisition equipment (MUSE-GE system) were collected to test model generalization between January 1, 1900, and December 31, 2023. Exactly 100 000 (from 65 291 unique patients) ECG recordings were selected from all consecutive ECGs with complete lead data and calculated measures. Patient characteristics of the Stanford Model Development and the Cedars-Sinai External Validation cohorts are provided in Table 1 (by patient) and in Table S2 (by ECG recording).A total of 1056 serial 12-lead ECGs from 22 healthy patients undergoing dofetilide loading, were each monitored for 24 hours after drug initiation.8 This was used specifically for the simulated continuous monitoring analysis. Device information for the ECGRDVQ data set was not specified.
    A separate retrospective clinical-encounter cohort of Stanford patients treated with dofetilide or sotalol, included 72 919 ECGs from 1676 unique patients across 2083 outpatient encounters (Stanford, 2008–2024). This cohort was used to assess the real-world clinical problem and validate 3DRECON-QT in the target population.
    The Medtronic CareLink database was accessed for all patients with an ICM implant (LINQ I and II, Medtronic, Minneapolis, MN) between 2016 and 2024. Only patients whose final transmission occurred during the period were accessible through the archives of older studies. Transmissions from a total of 387 patients were available. The timestamps from the presenting rhythm were compared with 12-lead ECGs acquired during routine clinical care. To ensure comparable physiological states between the 12-lead ECG and the ICM recordings, we included pairs obtained within ±7 days and with a heart rate difference |Δ heart rate |≤20 beats per minute. Twenty-six patients met these inclusion criteria.
    All ECG recordings used in the Stanford Model Development Cohort and the Real-World Class III Antiarrhythmic Cohort were acquired using the Philips iECG system. Cedars-Sinai external testing recordings were obtained through the MUSE-GE system. We extracted 10-s, 12-lead ECG segments and standardized all signals to a 500-Hz sampling rate. A 0-phase, fifth order Butterworth bandpass filter (0.5–40 Hz) was applied to remove baseline wander and high-frequency noise.
    In all analyses, the input to 3DRECON-QT is a single-lead, narrowly spaced chest wall signal. For model development, external testing, continuous monitoring, and real-world class III analyses, the input vector was a derived ICM signal calculated voltage differential between precordial leads V3 and V2. True ICM recordings were the input for the Stanford ICM Recording analysis. Twelve-lead ECGs were used as the target for reconstruction and the basis for gold standard QT labels. A summary of model input and 12-lead ECG use is provided in Table S1.
    We standardized each recording by subtracting the record-wide mean and dividing by its SD, so that all leads entered the network with zero mean and unit variance. We excluded recordings containing >70% 0-valued samples or any lead with both 0 mean and 0 SD. In addition, recordings lacking cardiologist-interpreted QT or RR annotations were removed. The CONSORT diagram for the development and external test sets is provided in Figure 2.
    Ground truth QT and RR interval measurements for our development and external validation cohorts were extracted from the clinical ECG systems after adjudication by cardiologists during routine clinical care. QTc values were calculated from the measured QT intervals and RR intervals using the Bazett formula.9 Across both data sets, a prolonged QT was defined according to the corrected QT interval and the QRS duration according to the dofetilide Food and Drug Administration drug insert labeling.10 For a narrow QRS (≤100 ms), a prolonged QTc was defined as >500 ms. For a wide QRS (>100 ms), a QTc >550 ms was determined to be prolonged.11 Based on a mixed-model pooled variance calculation (Detailed Methods), the interobserver variance (caused by reader bias) is 47.1 ms2 (SD, 6.86 ms) whereas the within-observer variance (caused by physiological variability, repeat measures, noise, etc) is 2392.7 ms2. This results in a variance ratio of 0.02, or a proportion of variance of 0.02, suggesting minimal effect of interobserver labeling on QT assessment and clinically robust QT labels (Figure S1).
    The 3DRECON-QT model is a deep learning architecture designed to reconstruct a full 12-lead ECG and estimate the QT interval from a single-lead input. It employs an encoder and decoder framework with a multitask learning objective, encouraging the network to learn physiologically and spatially meaningful latent features that jointly inform both waveform reconstruction and QT prediction (Figure 2).
    A 10-s single-lead ECG signal with 500-Hz sampling is processed through a squeeze-and-excitation ResNeXt encoder, which extracts a compact latent representation of cardiac electrical activity.12,13 Two task-specific “heads” operate in parallel: a reconstruction head, which generates the 12-lead ECG by conditioning on the spatial coordinates (θ and ϕ) of each lead,14 and a QT prediction head, which applies a transformer decoder to the shared latent representation to estimate the QT interval.15 Model optimization used a weighted composite loss across the 2 tasks, stochastic gradient descent with cosine annealing, and early stopping based on validation performance.
    Implementation was performed in Python (PyTorch v2.3.1).16 Detailed architectural, training, and hyperparameter details are provided in the Supplemental Methods.
    To assess the performance of 3DRECON-QT to detect changes in QT interval during continuous monitoring, we validated the model in the Continuous QT Monitoring Cohort (ECGRDVQ PhysioNet Dataset), which investigates the pharmacokinetic profile of dofetilide taken with oral ingestion during safety studies of the medication.8 We assessed 3DRECON-QT for detecting drug-induced QT prolongation using 24-hour ECGs from the ECGRDVQ dofetilide-loading data set (simulating ambulatory monitoring). The input signal is the same as described in the “ECG Recording and Preprocessing” section. We examined the temporal trends of aggregated mean predicted and ground truth QT/QTc. We performed classification analysis for prolonged QT interval, using the thresholds as described above for dofetilide loading and, secondarily, a ≥15% increase in QTc from baseline.17
    To assess the clinical impact of an enhanced monitoring strategy in a high-risk patient population, we conducted a retrospective encounter-based analysis using the Stanford Real-World Class III Antiarrhythmic Exposure Cohort, a data set comprising 1676 unique patients receiving outpatient dofetilide or sotalol treatment at Stanford between 2008 and 2024 at the time of a hospital or emergency room encounter for any reason. We categorized the primary diagnosis from the encounter using the Clinical Classifications Software Refined for International Classification of Diseases, Tenth Revision, clinical modification codes to express which types of clinical encounters were associated with prolonged QT events on the first ECG of each encounter.18 We then established a case-control framework for serious arrhythmic events (torsades de pointes, ventricular fibrillation, and sudden cardiac death), calculating adjusted odds ratios (ORs) using the Fisher exact test. Last, we ensured that QT assessments in this high-risk population were appropriately classified using 3DRECON-QT performance on this clinical data set using classification analysis and receiver operator characteristics.
    We evaluated 3DRECON-QT using Medtronic LINQ ICM signals paired with a clinical 12-lead ECG obtained within ±7 days. Analyses were performed at the patient level (one paired QTc value per patient). For each patient, 3DRECON-QT generated predictions from the real ICM recordings, with the paired 12-lead ECG providing the reference QTc value. We repeated the QTc analysis using the derived ICM signal from the paired 12-lead ECG. Agreement with the reference QTc across patients was assessed using Pearson correlation coefficient and the mean absolute error (MAE) between the prediction and ground truth. We also compared waveform morphology between the real and simulated ICM signals and between the corresponding, model-predicted 12-lead reconstructions. R-peaks were detected using the Neurokit2 Python package, after which individual beats were segmented, temporally aligned to the R-peak, and normalized before overlay.19
    Statistical analysis was performed in Python (v3.8) using SciPy and Scikit-learn.20 Descriptive statistics were reported as mean±SD or median [interquartile range]. For QT prediction performance, agreement between predicted and ground truth values was assessed using Bland-Altman analysis, Pearson correlation coefficient (r), and linear regression to calculate R2, and MAE. Classification performance for prolonged QTc detection was evaluated using sensitivity, specificity, positive predictive value, negative predictive value, F1 score, area under the receiver operating characteristics curve (AUROC), and area under the precision-recall curve (AUPRC). CIs (95% CI) were generated using nonparametric bootstraps (1000 iterations, random sampling with replacement at the ECG level). Sensitivity analysis used patient-level clustered bootstrap with bias-corrected and accelerated intervals to account for multiple ECGs per patient. Association between QTc prolongation and clinical events was assessed using the Fisher exact test for univariate analysis and multivariable logistic regression to calculate adjusted ORs controlling for age, heart failure, atrial fibrillation, drug class, and laboratory values (serum creatinine, magnesium, and potassium). P values <0.05 were considered statistically significant.

    RESULTS

    There were no significant differences between patient characteristics across the internal splits of the Stanford Model Development Cohort (aggregate demographics: mean age, 61.9 years; 52.2% male; race or ethnicity, White 57.6%, Asian 13.4%, Black 5.7%, Hispanic or Latino 12.7%). The Cedars-Sinai External Validation Cohort contained older individuals with a higher proportion of Black individuals compared with the Stanford Model Development Cohort (Table 1).
    Prolonged-QT classification using derived ICM recordings achieved an AUROC of 0.942 (95% CI, 0.941–0.943), AUPRC of 0.632 (95% CI, 0.626–0.639), and an accuracy of 0.957 (95% CI, 0.955–0.957) in the test partition of the Stanford Model Development Cohort. In the Cedars-Sinai External Validation Cohort, performance was similarly high, with an AUROC of 0.943 (95% CI, 0.941–0.945), AUPRC of 0.781 (95% CI, 0.774–0.789), and an accuracy of 0.925 (95% CI, 0.923–0.926) (Figure 3A).
    The median ground truth QTc was 411 ms (interquartile range, 391–441 ms). In the test split of the Stanford Model Development Cohort, the model-predicted median QTc was 418 ms (interquartile range, 401–441 ms). The Bland-Altman analysis for QTc revealed a bias of –5.46±29.38 ms (Figure 3C).
    To qualify 12-lead ECG reconstructions systematically, we analyzed performance across the Stanford Model Development test set, achieving mean Pearson correlation of 0.70 across all 12 leads (Table S3). Saliency analysis of the output signals shows activation across leads at clinically expected time points, including QRS onset and T wave offset along with accurate 12-lead reconstructions (Figure S2). To evaluate the contribution of the spatial encoding to QT prediction, we compared 3DRECON-QT with an ablated variant without the theta (spatial angle) encoder. The model with spatial encoding achieved QTc MAE of 25.9 ms and Pearson r of 0.73, whereas the ablated model without spatial encodings showed markedly degraded performance (QTc MAE, 39.23 ms; Pearson r, 0.04), despite similar training loss convergence (Figure S3).
    A linear regression of the predicted QT (uncorrected) resulted in an R2 value of 0.747 (95% CI, 0.742–0.752) and a Pearson correlation coefficient of 0.864 (95% CI, 0.861–0.867; Figure 3B). This represented a MAE of 17.52 ms (95% CI, 17.38–17.64) for the predicted QT interval in the test split of the Stanford Model Development Cohort.
    In the Cedars-Sinai External Validation Cohort, the MAE for QT prediction was 21.14 ms (95% CI, 20.99–21.29). The prolonged QTc classification revealed an AUROC of 0.943 (95% CI, 0.941–0.945) and an AUPRC of 0.781 (95% CI, 0.774–0.789). See Table 2 for full test characteristics.
    A sensitivity analysis using patient-level clustered bootstraps using bias-corrected and accelerated intervals is presented in Table S4 and showed minimal change from standard bootstrapping.
    In patients from the Continuous QT Monitoring Cohort (ECGRDVQ PhysioNet Dataset), 3DRECON-QT demonstrated a statistically significant correlation between predicted QT intervals and ground truth values (Pearson r=0.851; P<0.001; Figure 4). For serial QT measurements over the duration of recording, the MAE was 17.8 ms for QT interval. For detecting prolonged QTc using the absolute risk threshold (see Methods, “QT Annotation”), the model achieved an AUROC of 0.936 (95% CI, 0.900–0.967) and AUPRC of 0.395. For the secondary task of detecting a ≥15% increase in QTc from baseline, the model achieved an AUROC of 0.816 (95% CI, 0.790–0.843) and AUPRC of 0.538.
    In the Stanford ICM Recording Cohort, 26 patients met criteria for time Δ (±7 days) between the ICM recording and the clinical reference 12-lead ECG. Figure 5A shows beat-wise morphological comparison between an actual ICM recording and simulated ICM from the 12-lead ECG. 3DRECON-QT predicted QTc values from the ICM recording with a Pearson correlation of r=0.824 (95% CI, 0.67–0.92) and mean absolute error of 17.5 ms (95% CI, 12.4–23.2) with the gold standard QTc (Figure 5B). A comparison for the same patient for the ICM and 12-lead reconstructions is shown in Figure 5C. Using derived ICM signals from the reference 12-lead ECG for the same cohort had a Pearson r=0.937 (95% CI, 0.885–0.970) and mean absolute error of 18.6 ms (95% CI, 14.1–23.6) with the gold standard QTc measurement. The difference in mean absolute error between the 2 methods was –1.1 ms (95% CI, –5.5 to 3.5).
    Baseline characteristics for patients in the Stanford Real-World Class III Antiarrhythmic Cohort are provided in Table S5. In this cohort, 1 in 6 patients (16.5%, 277 of 1676) had at least one documented episode of high-risk QTc prolongation after their initial inpatient loading on presentation to medical care for any indication. These events occurred during outpatient visits for a wide range of primary diagnoses, not just arrhythmia management follow-up. Figure 6A shows that approximately one-third of admissions with prolonged QT were for noncardiac causes, whereas one-third of cases occurred during a primary visit for arrhythmia evaluation, and one-third for non-arrhythmic cardiac evaluation.
    The overall incidence of ventricular arrhythmic event–related hospitalization (torsades de points, ventricular fibrillation, and sudden cardiac death) in the cohort was 1.37%. This was increased for patients with prolonged QTc compared with those without (3.97% versus 0.86%; univariate OR, 4.78; P<0.05). After adjusting for age, heart failure, atrial fibrillation, drug class (sotalol/dofetilide), and laboratory values including creatinine, potassium, and magnesium, the risk of event remained 4-fold (OR, 4.24; 95% CI, 1.81–9.90; P<0.05).
    For patients with an active outpatient dofetilide or sotalol prescription, 3DRECON-QT accurately identified prolonged QTc based on the derived ICM signal with an AUROC of 0.94 and an F1 score of 0.60 (Figure 6B).

    DISCUSSION

    3DRECON-QT accurately predicts QT/QTc within clinical tolerances (mean error, 18–21 ms) with AUROC 0.94 for classifying prolonged QT from a single-lead, nonstandard ECG signal. The system was robust to patient populations across 2 health systems with varying demographics. Further, when applied to a continuous monitoring scenario during pharmacological therapy, the model predicted QTc changes subsequently validated by observed serial measurements. Moreover, the model revealed QT prolongation in patients evaluated for noncardiac conditions, which could easily be missed in routine practice.
    In our real-world analysis, 1 in 6 patients receiving class III antiarrhythmic medications presented from the ambulatory setting with dangerous QTc prolongation for encounters related to both cardiac and noncardiac causes. These episodes confer a >4-fold increase in serious ventricular arrhythmia events despite previous successful in-hospital drug loading. Overall event rates of ventricular arrhythmia in patients with and without prolonged QTc were 0.86% and 3.97%, respectively, which is consistent with rates reported in randomized trials21,22 and systematic review.4 The causal relationship between QTc prolongation and arrhythmia events was not adjudicated in this study. Out-of-hospital arrests that did not reach our center would be underreported by this analysis. This analysis underscores the unmet clinical need of safe ambulatory monitoring of these medications.
    Previous evidence that successful loading at a given dose of antiarrhythmic dose does not guarantee safety of that dose at future encounters.23 Current guidelines require a 12-lead ECG for QT assessment at least every 3 to 6 months for patients on class III antiarrhythmic agents.5,24 However, a recent statement from the Heart Rhythm Society highlights that although single-lead wearable devices can provide serial QT interval measurements, their accuracy and clinical utility remain inferior to standard 12-lead ECGs.25 Our data show how 3DRECON-QT may address this gap by providing near real-time QT safety monitoring from existing device infrastructure without additional leads, patient visits, or patient compliance burden.26 Cloud-based artificial intelligence pipelines are already in use (such as AccuRhythm AI, Medtronic, Minneapolis, MN) to filter out false-positive events in rhythm detection upstream of the provider portal, which provides a vantage point for immediate implementation. Last, this paradigm could possibly shorten or obviate mandatory hospital initiation of these medications for low-risk patients through hybrid inpatient-outpatient safety protocols and by providing objective triggers for dose reduction, electrolyte assessment, or drug discontinuation.27
    The multitask encoder in 3DRECON-QT learns latent representations that preserve ventricular repolarization vectors, explaining high-fidelity QT inference despite limited spatial input. Previous deep-learning approaches have estimated QT from lead I or smartwatch tracings but lacked external validation.28 Further, the more fixed morphology of the lead I recording is less generalizable to insertable or chest-worn recorders that have variable and nonstandard recording vectors. To achieve long-term, continuous monitoring, a more flexible model is required. In addition, the training data involve the largest QT assessment study to date with strict patient-level splits, which contributed to its robust performance in multi-institution external testing.
    Limitations of this study include its retrospective design and the fact that QT labels are derived from clinician-reviewed ECGs during routine clinical care rather than those specifically ordered for QT assessment. Simultaneous ICM and 12-lead ECG recordings are not available because of the proprietary nature of the ICM recordings. Our method of training 3DRECON by deriving vectors approximating the ICM allows us to overcome this limitation. Physiological and positional factors such as circadian variation, posture, and rate dependency are known to influence QT measurements and may affect the relationship between the ICM and 12-lead ECG signals. The broad inclusion of patient states in the development and external testing sites supports the generalizability of the model. The high correlation of the QTc predictions from the true ICM recordings in our data set supports the use of the V3-V2 vector surrogate for training and the generalizability of the 3DRECON-QT network. Prospective simultaneous collection of 12-lead signals with digital ICM recordings would provide further clinical validation and the opportunity to improve model performance through fine tuning. Twelve-lead reconstructions are only validated for measurement of the QT interval. Last, the clinical impact for patients receiving other QT-prolonging agents such as antipsychotics, antiemetics, antiepileptics, and others is not assessed in this study and may have a different benefit threshold compared with class III antiarrhythmics to deploy this monitoring strategy.
    Future work may involve deployment in patient populations with preexisting ICMs or chest-worn patches or consumer devices to detect safety concerns. In our high-risk cohort, precision exceeding 75% indicates a modest false-positive burden, with fewer than 2 notifications needed to identify one episode of prolonged QTc event. Notifications for prolongation events may be resolved by a 12-lead ECG and medication review. Conversely, the sensitivity of >80% from a single measurement may offer a substantial improvement in safety over current sporadic clinic-based monitoring. Although continuous monitoring introduces additional challenges such as motion artifacts, future deployment may incorporate noise-filtering strategies. Studying appropriate patient selection is critical to increase drug safety without creating undue health care use.

    Conclusions

    3DRECON-QT demonstrates that deep-learning reconstruction of latent spatial information from single-lead signals can meet guideline-level QT accuracy. This model may address a substantial hidden burden of high-risk prolongation after class III antiarrhythmic initiation and other QT-prolonging drugs.

    Supplementary Material

    Supplemental Material is available at https://www.ahajournals.org/doi/suppl/10.1161/CIRCULATIONAHA.125.077494.
    Checklists
    Detailed Methods
    Tables S1–S5
    Figures S1–S3
    PWDr. Paul Wang

    Dr. Paul Wang

    Cardiology · Cardiac Electrophysiology
    Stanford Medicine
    paulwang