Journal of electrocardiology · 2025
Are single-lead consumer ECG devices enough for accurate QTc monitoring? A calibrated uncertainty AI approach.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Doggart Peter, Bond Raymond, Finlay Dewar, Guldenring Daniel, Albert Dave, Kennedy Alan
Аннотация
BACKGROUND: While single‑lead ECGs offer accessibility, their performance and reliability for QTc assessment remains uncertain. Current State-of-the-art AI systems, though promising, often lack transparency, raising concerns about clinical trustworthiness.
METHODS: We developed an uncertainty-aware AI model to measure RR/QT intervals from single‑lead ECGs. Training used retrospective datasets (∼400,000 ECGs) from clinical 12‑lead systems, with testing on 2050 manually annotated AliveCor KardiaMobile single‑lead (lead-I) ECGs. QTc Fridericia values were calculated using a first-order approximation and delta method for uncertainty propagation. Performance was evaluated using (1) Bland-Altman analysis and uncertainty calibration (ENCE), (2) Classification performance for QTc prolongation at 450/470/500 ms thresholds using probabilistic outputs and area under the receiver operator curve (AUROC) and (3) QT nomograms for risk visualization.
RESULTS: The model demonstrated strong QTc estimation performance with integrated uncertainty quantification. After excluding 11 % of ECGs flagged as unreliable through uncertainty analysis, the remaining 89 % showed excellent agreement with 12‑lead reference values (mean bias: -0.7 ms, limits of agreement: -42 to +41 ms). The excluded high-uncertainty cases exhibited substantially larger errors (mean bias: 15.3 ms, limits of agreement: -104 to +134 ms). Uncertainty estimates were well-calibrated (ENCE: 14.6 %), enabling effective identification of unreliable predictions for expert review while maintaining clinical-grade accuracy in retained cases. Classification achieved excellent AUROCs in low-uncertainty cases: 0.87 [95 % CI: 0.85-0.89](≥450 ms), 0.90 [95 % CI: 0.86-0.93](≥470 ms), 0.96 [95 % CI: 0.92-1.00](≥500 ms).
CONCLUSIONS: By integrating uncertainty quantification, this AI model enables trustworthy QTc monitoring using single‑lead ECGs, dynamically flagging unreliable measurements. This approach enhances clinical safety and expands access to cardiac risk assessment in remote settings. The methodology sets a precedent for developing transparent, reliable AI tools in healthcare, prioritizing trust over traditional explainability.
METHODS: We developed an uncertainty-aware AI model to measure RR/QT intervals from single‑lead ECGs. Training used retrospective datasets (∼400,000 ECGs) from clinical 12‑lead systems, with testing on 2050 manually annotated AliveCor KardiaMobile single‑lead (lead-I) ECGs. QTc Fridericia values were calculated using a first-order approximation and delta method for uncertainty propagation. Performance was evaluated using (1) Bland-Altman analysis and uncertainty calibration (ENCE), (2) Classification performance for QTc prolongation at 450/470/500 ms thresholds using probabilistic outputs and area under the receiver operator curve (AUROC) and (3) QT nomograms for risk visualization.
RESULTS: The model demonstrated strong QTc estimation performance with integrated uncertainty quantification. After excluding 11 % of ECGs flagged as unreliable through uncertainty analysis, the remaining 89 % showed excellent agreement with 12‑lead reference values (mean bias: -0.7 ms, limits of agreement: -42 to +41 ms). The excluded high-uncertainty cases exhibited substantially larger errors (mean bias: 15.3 ms, limits of agreement: -104 to +134 ms). Uncertainty estimates were well-calibrated (ENCE: 14.6 %), enabling effective identification of unreliable predictions for expert review while maintaining clinical-grade accuracy in retained cases. Classification achieved excellent AUROCs in low-uncertainty cases: 0.87 [95 % CI: 0.85-0.89](≥450 ms), 0.90 [95 % CI: 0.86-0.93](≥470 ms), 0.96 [95 % CI: 0.92-1.00](≥500 ms).
CONCLUSIONS: By integrating uncertainty quantification, this AI model enables trustworthy QTc monitoring using single‑lead ECGs, dynamically flagging unreliable measurements. This approach enhances clinical safety and expands access to cardiac risk assessment in remote settings. The methodology sets a precedent for developing transparent, reliable AI tools in healthcare, prioritizing trust over traditional explainability.