This study introduces a lightweight hybrid framework that integrates Convolutional Neural Networks (CNN) and Fast Fourier Transform (FFT) methodologies aimed at enhancing the detection of cardiac abnormalities through electrocardiogram (ECG) analysis. Given the vital role of ECGs as a non-invasive tool for monitoring heart health, this framework is particularly beneficial in scenarios where specialized expertise may not be readily accessible, especially within resource-limited clinical environments. The proposed architecture adeptly merges time-domain morphological representations obtained from raw ECG signals with physiologically pertinent spectral features, facilitating accurate and efficient classification of cardiac conditions.
The study undergoes assessment utilizing available data from the PTB-XL dataset, implementing a rigorous 10-fold cross-validation approach for both binary and five-class multi-label classification tasks. Notably, in addition to the conventional discrimination metrics commonly employed, the researchers prioritize model reliability through techniques such as probability calibration and implement a stringent patient-wise validation method. The performance metrics reveal that the model achieved an accuracy of 92.42% for binary classification and an area under the curve (AUC) score of 97.8%. Moreover, for the five-class multi-label classification, the model attained a macro-AUC of 92.46%. Calibration assessments indicate well-calibrated probabilities with a minimal Expected Calibration Error (ECE).
In summary, this framework delivers competitive performance while maintaining efficiency, characterized by a mere 87,000 parameters and an operational load of 0.26 GFLOPs. These findings underscore the promising capabilities of hybrid lightweight architectures alongside calibration-aware evaluations to support reliable AI-assisted diagnostics in ECG interpretation, particularly within resource-constrained healthcare settings.
Cardiovascular diseases (CVDs) represent a significant global health crisis, contributing to approximately 32% of worldwide mortality. Timely intervention and accurate diagnosis are pivotal in mitigating morbidity linked to cardiac disorders. While electrocardiograms (ECGs) serve as a crucial diagnostic tool for assessing the heart's electrical activities, their interpretation often necessitates extensive clinical expertise, which can be lacking in low- and middle-income nations due to a shortage of trained cardiologists. This gap frequently results in delayed diagnostics and treatment, especially in rural or resource-constrained healthcare environments. Hence, developing effective automated ECG analysis systems is essential to bolster clinical decision-making abilities and facilitate early cardiology screening.
Recent advancements in deep learning have opened the door for automating ECG-based diagnostics, which have shown considerable promise in improving early diagnosis and patient triage processes. Notable architectures, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and hybrid models, have showcased substantial performance enhancements, achieving results equivalent to expert cardiologists in certain contexts. However, the reliability and generalizability of these systems remain areas of ongoing scrutiny. Many previous approaches utilize proprietary datasets, which restrict reproducibility, foster potential data leakage, and fail to assess critical clinical metrics, such as model calibration.
Several studies have illustrated the efficacy of employing CNN architectures for ECG classification. For instance, one highlighted CNNs that work directly on raw signals, achieving significant accuracy and AUC scores, while others have integrated unique mechanisms to refine ECG feature capture. However, issues with reproducibility and model reliability still linger, necessitating the search for frameworks capable of delivering reliable and interpretable ECG assessments for clinical use.
The proposed hybrid framework combines both CNN and FFT methodologies. The model architecture is designed to optimize performance while maintaining a low computational footprint to enhance its applicability in resource-constrained environments. Its evaluation incorporates an official 10-fold cross-validation protocol on the PTB-XL dataset, ensuring robustness in performance verification through diversified testing.
Additionally, model assessment encompasses both traditional classification metrics and calibration assessments, thus providing a comprehensive evaluation of its effectiveness. In this regard, emphasis on calibration improves the model's applicability in real-world scenarios where accurate probability estimates are paramount.
The study's findings highlight an accuracy of 92.42% and an AUC of 97.8% for binary classification tasks, while satisfying macro-AUC metrics of 92.46% for multi-class classifications. Calibration analysis indicates the model's probability estimates are well-calibrated, with low Expected Calibration Error (ECE) values. This affirms the functionality of the hybrid framework not just in accuracy but also in reliability—key factors that enhance user trust in AI-assisted diagnostic tools.
The research provides new insights into the efficacy of hybrid models for ECG classification, harnessing both CNN and FFT capabilities. By attaining competitive performance metrics alongside demonstrating efficient model operation, the study signals considerable promise for integrating such technologies into broader healthcare applications, especially in settings reliant on automated diagnostics. The incorporation of calibration metrics further aligns clinical evaluations with practical applications, enabling greater confidence in deploying such systems for patient care.
This research underscores the potential of leveraging hybrid frameworks that integrate deep learning components with statistical methods like FFT for cardio-related diagnostic tasks. The positive outcomes reveal pathways for effective, automated ECG analyses, crucial for enhancing healthcare delivery amidst global disparities in cardiac care accessibility.