What does this research mean for the field?

The deep neural network (DNN) achieves high precision in detecting normal ECGs, with an AUC of 0.933 in the retrospective analysis and 0.918 in the prospective analysis. Novelty: ClaimNovelty.CONFIRMATORY. Consensus alignment: ConsensusAlignment.SUPPORTS_CONSENSUS.

What question did this study set out to answer?

This research focuses on developing an AI-based system to classify normal ECGs automatically, enhancing diagnostic workflows.

February 8, 2026

Artificial intelligence in the electrocardiogram: automatic diagnosis of the normal ECG in a Tele-electrocardiogram service

Resultado clave

The deep neural network achieved high precision (0.968) and AUC of 0.918 for detecting normal ECGs in a telehealth system, aiding diagnostic efficiency.

Puntos clave

This research focuses on developing an AI-based system to classify normal ECGs automatically, enhancing diagnostic workflows.
Developed a deep neural network (DNN) for ECG classification, trained on a dataset of 2,933,600 ECGs.
Preprocessed ECG leads including filtering and zero-padding to standardize data for analysis.
Conducted both retrospective and prospective testing phases to evaluate DNN accuracy against cardiology reviews.
In retrospective analysis, the DNN achieved an AUC of 0.933; prospective analysis yielded an AUC of 0.918.
Precision of 0.968 and diverse recall metrics for different thresholds were noted, with the F1-optimized threshold showing promising performance.
The DNN supports efficient processing of normal ECGs, enabling better resource allocation for specialists.

PICO estructurado

Does a deep neural network accurately detect normal ECGs compared to cardiologist evaluation in a large-scale tele-electrocardiogram service?

Población

Dataset of 2,933,600 ECGs for training/validation, 8,933 ECGs for retrospective testing, and 10,369 ECGs for prospective testing from the Telehealth Network of Minas Gerais (TNMG) supporting the Brazilian Public Health System.

Intervención

Deep neural network (DNN) with a convolutional architecture and five residual blocks for automated detection of normal ECGs.

Comparador

Cardiologist evaluation (reference standard).

Resultado

Detection of normal ECGs, measured by Area Under the Receiver-Operator Curve (AUC), precision, recall, specificity, and Negative Predictive Value (NPV).surrogate

A deep neural network can accurately identify normal ECGs in a large-scale telehealth system, which could improve diagnostic efficiency by allowing specialists to focus on abnormal cases.

Resultado numérico

Tasa de eventos absoluta: 0% vs 0%

Resumen

Abstract Background The Telehealth Network of Minas Gerais (TNMG) provides tele-electrocardiography (tele-ECG) services to support the Brazilian Public Health System (SUS). It analyzes around 7,000 ECGs daily with prospects for expansion. The large volume of ECGs can overwhelm specialists’ availability, compromising diagnostic accuracy. Since most ECGs are normal, an artificial intelligence (AI) automated classification system could improve diagnostic efficiency, allowing specialists to focus on abnormal cases. This study aims to develop such a system to improve the efficiency and speed of cardiac diagnostics, with potential for commercialization and expansion into traditional ECG systems, cloud services, and international markets. Methods We developed a deep neural network (DNN) to detect normal ECG using a dataset of 2,933,600 ECGs (81% training, 19% validation) evaluated by TNMG cardiologists. Preprocessing for the 8 essential ECG leads (I, II, V1-V6) includes resampling to 400Hz, filtering low and powerline frequencies, and zero-padding or truncating traces to 4096 samples per lead. The DNN has a convolutional architecture with five residual blocks and a fully connected layer with sigmoid activation, yielding probability scores for ECGs. In the retrospective phase, the test set includes 8,933 ECGs with at least 2 concordant cardiology reviews. In the prospective phase, the DNN was integrated into TNMG’s real-time platform and further tested on 10,369 ECGs reported by cardiologists who could not access the model’s results. An ECG is classified as normal if its score exceeds a threshold. We consider two scenarios: one optimized to maintain =95% precision with maximum recall, preventing telehealth system overload (referred to as precision-optimized threshold), and another maximizing the F1 score to balance precision and recall (referred to as F1-optimized threshold). These thresholds are derived from the retrospective phase and tested in the prospective phase. Results In the retrospective analysis, the DNN achieved an Area Under the Receiver-Operator Curve (AUC) of 0.933, while the prospective analysis reached an AUC of 0.918. Precision, recall, specificity, and Negative Predictive Value (NPV) metrics for the precision-optimized (0.968) and F1-optimized (0.483) thresholds are shown with the Precision-Recall Curves in Figure 1. Conclusions: The study shows that the DNN achieves high-precision detection of normal ECGs with reasonable recall for the precision-optimized threshold and promising metrics (mean 0.8) for the F1-optimized threshold in both analyses. This supports its integration into TNMG's tele-ECG system, reaching over 1,400 municipalities through SUS, with potential for cost savings, improved healthcare, and global expansion.Precision-Recall Curves

Me gusta

Guardar

Cite This Study

Paixao et al. (Sat,) reported a other. The deep neural network achieved high precision (0.968) and AUC of 0.918 for detecting normal ECGs in a telehealth system, aiding diagnostic efficiency.

synapsesocial.com/papers/698827f00fc35cd7a8846ef0 https://doi.org/https://doi.org/10.1093/eurheartj/ehaf784.4407

Me gusta

Guardar