False alarms in medical equipment, particularly in intensive care units (ICUs), lead to decreased trust and potentially missed real events. This study focuses on reducing false positive alarms in heartbeat segmentation, a crucial step in ECG signal processing. Existing approaches often concentrate on classification stages, neglecting errors propagated from the segmentation stage. This work proposes a novel method for improving heartbeat segmentation accuracy by employing deep learning techniques to validate the output of a third-party QRS complex detection algorithm. The use of CNNs allows for real-time processing and easy integration with dedicated hardware, improving the reliability of medical equipment. The proposed method leverages the advantages of CNNs for pattern recognition, offering a robust and potentially deployable solution to a critical problem in medical signal processing. The study uses two well-established databases, MIT-BIH and CYBHI, to evaluate the performance of the proposed approach, comparing it against a standard Pan-Tompkins algorithm. The primary goal is to demonstrate the effectiveness of CNNs in enhancing the accuracy of heartbeat segmentation while maintaining real-time feasibility.

The problem of excessive false alarms in medical equipment has been extensively researched. Many existing approaches rely on ECG signals, which provide information on the heart's electrical activity. The 2015 PhysioNet/CinC Challenge highlighted the importance of reducing false alarm rates, particularly for life-threatening arrhythmias. Plesinger et al. achieved an impressive score by using a multi-channel approach with filtered signals, spectral features, and heuristic rules. However, most studies focus on reducing false alarms in the classification stage, neglecting errors stemming from the segmentation stage. Some promising approaches include signal quality assessment, like that of Behar et al., which uses machine learning to distinguish between good and bad signal quality. Other researchers have utilized multi-modal approaches, combining ECG data with other physiological signals such as blood pressure. This study differs by proposing a deep learning-based approach that directly validates the QRS complex patterns detected by a third-party algorithm, focusing on the shape of the heartbeat rather than overall signal quality. This method can benefit from hardware acceleration, allowing for easy integration into real-world medical equipment and potential for online learning improvements.

The proposed method comprises six steps: database split, pre-processing, CNN training, R-peak detection, R-peak validation, and evaluation. The databases (MIT-BIH and CYBHI) are split into training and testing sets. Pre-processing involves segmenting the signals into fixed-length windows (300 samples for MIT-BIH, 833 for CYBHI) and applying data augmentation techniques. Data augmentation for positive samples includes shifting the R-peak and attenuating P and T waves to enhance model robustness. Negative samples are generated from segments between R-peaks. The Pan-Tompkins algorithm is used as the third-party R-peak detector. A seven-layer CNN (architecture detailed in the paper) is trained for binary classification (heartbeat/no heartbeat). The CNN validates R-peaks detected by Pan-Tompkins; if the CNN agrees, the R-peak is considered valid. The method is evaluated using sensitivity, positive predictive value (+P), and F-score. The study also explores the computational cost of the CNN inference on different hardware (CPU, GPU, NVIDIA Jetson Nano).

The proposed CNN-based validation method significantly improves the positive predictive value (+P) of the Pan-Tompkins algorithm while maintaining a comparable F-score. On the MIT-BIH database, +P increased from 97.84% to 100.00%, with a slight decrease in sensitivity (95.79% to 92.98%). On the CYBHI database, +P increased from 90.28% to 96.77%, with a minor decrease in sensitivity (96.95% to 95.71%). The CNN model demonstrates good generalization across both databases. Computational cost analysis shows that the CNN inference is fast enough for real-time applications, even on resource-constrained hardware like the NVIDIA Jetson Nano. Analysis of the CNN filter outputs reveals differences between the models trained on MIT-BIH and CYBHI, potentially reflecting the different noise levels in the datasets. The higher noise in CYBHI may account for the slight decrease in sensitivity observed in the results. Examples illustrate cases where the CNN correctly rejects false positives and, conversely, misclassifies true positives. This trade-off between positive predictive value and sensitivity is discussed, suggesting that this approach prioritizes reducing false alarms.

The results demonstrate that integrating a CNN as a validation step significantly improves the positive predictive value of a widely-used heartbeat segmentation algorithm. This is crucial for medical applications where false positive alarms can have significant consequences. While a slight decrease in sensitivity is observed, the overall F-score remains largely unaffected or even improves in one database, suggesting a worthwhile trade-off in many contexts. The successful application across two distinct datasets with varying noise characteristics underlines the robustness and generalizability of the proposed method. The real-time feasibility of the approach, especially with hardware acceleration, opens the way for practical implementation in medical equipment. The findings highlight the potential for deep learning methods to enhance the accuracy and reliability of existing medical signal processing algorithms.

This work presents a novel deep learning-based method for enhancing heartbeat segmentation accuracy. The results demonstrate that a CNN can effectively validate R-peaks detected by the Pan-Tompkins algorithm, increasing positive predictive value significantly while maintaining a comparable F-score. The method's real-time feasibility on resource-constrained hardware suggests its potential for integration into real-world medical devices. Future work could explore the application of noise reduction techniques, the use of pre-trained models, and the ability to classify diverse heartbeat patterns beyond the normal heartbeat. The availability of the code and data allows other researchers to build upon these results.

The study uses only two databases, which might limit the generalizability of the findings. The sensitivity of the method is slightly reduced compared to the baseline Pan-Tompkins algorithm, which may be a concern in applications requiring high sensitivity. The model is trained to recognize the pattern of normal heartbeats; its performance on arrhythmic beats requires further investigation. The data augmentation strategy might not fully capture the variability of real-world ECG signals. Finally, the specific implementation of the Pan-Tompkins algorithm used as the baseline could influence the results, as different implementations might yield different performance levels.