Design and Development of Atrial Lead System with Integrated Artificial Intelligence Models for Enhanced Diagnosis of Atrial Arrhythmias

Abstract

The electrocardiogram or ECG remains the most widely used and cost-effective tool for diagnosing cardiac arrhythmias. The morphological features of the P-wave on an ECG offer valuable insights into abnormalities in interatrial and atrioventricular (AV) conduction. Analyzing changes in P wave morphology is essential for identifying and characterizing atrial arrhythmias. Despite the extensive use of ECG, discriminating between different types of atrial arrhythmias can be time-consuming and prone to false positives, primarily due to the small size of P-waves and their vulnerability to interference. Traditional 12-lead ECG systems often struggle to detect these subtle P-wave abnormalities, leading to diagnostic redundancies and increased risk of misdiagnosis. Atrial arrhythmias, such as Atrial Fibrillation (AF), Atrial Flutter (AFL), and Atrial Tachycardia (AT), are common among hospitalized individuals with cardiac abnormalities and can significantly impact morbidity and mortality rates. Therefore, improving the quality of ECG signals is essential for minimizing false positives and enhancing diagnostic accuracy in the identification of atrial arrhythmias. The research reported in this thesis was performed to enhance the detection of atrial arrhythmias by improving the signal strength of atrial activity (P/f/F -waves) through optimal lead modification and the implementation of automated algorithms. Machine Learning (ML) and Deep Learning (DL) based algorithms driven by Artificial Intelligence (AI) were used to optimize lead selection and accurately classify arrhythmias, thereby facilitating improved detection and diagnosis of atrial arrhythmias. In the initial stage of the study, a novel Atrial Lead System (ALS) was introduced to enhance the strength of P-wave signals. Advanced Gradient Boosting (GB) and DL algorithms were further employed to improve the detection of atrial activity by ranking optimal bipolar leads. Various ML models, including GB and DL algorithms, were used to evaluate and rank optimal bipolar leads based on P-wave parameters, indices, and AV ratios. AL-I and AL-II were found to have significantly higher median amplitudes, RMS values, and area under the curve for recorded P-waves compared to other leads. The proposed models identified P-lead, AL-II, and AL-I as the top three leads. The selection of optimal leads is critical for improving the detection of P-wave changes. Our study proposed an automated lead selection technique using the CatBoost ML model to enhance the detection of P-wave changes among optimal bipolar leads under various heart rates (HR). P-wave features and AV ratios were extracted for statistical analysis and ML classification. CatBoost outperformed other ML models in Standard Limb Lead-II (SLL-II), achieving the highest accuracy and sensitivity. It demonstrated superior performance for AL-I and AL-II compared to other optimal leads, identifying them as the top two best-performing optimal leads for enhanced detection of P-wave alterations. This study also explored ALS to improve signal strength, noise resistance and reliability for smart and mobile health (mHealth) applications. ECG data from individuals with normal Sinus Rhythm (SR) and various atrial arrhythmias, including AF, AFL, and Left Atrial Enlargement (LAE), were recorded using ALS leads (AL-I and AL-II). Statistical analysis, correlation analysis, and reliability analysis using Bland-Altman plots confirmed significant differences in amplitude, duration, area, and A/V ratio between SR and AF populations when comparing SLL- II, AL-I, and AL-II. The ALS leads demonstrated a substantial improvement in signal quality, Signal-to-Noise Ratio (SNR) and reliability when compared to SLL-II for both SR and AF conditions. Finally, an automatic classification technique using a 1D-CNN and BiLSTM model ensemble was proposed for differentiating atrial arrhythmias from normal SR. Dataset preparation included data from Lead-II obtained from Chapman University and Shaoxing People’s Hospital (CUSPH), which underwent preprocessing, segmentation, and augmentation to ensure balanced classes. The proposed model achieved the highest accuracy of 94% across cross-validation and testing datasets when classifying atrial arrhythmias, showcasing its effectiveness in detecting various atrial arrhythmias. These findings highlight the effectiveness of ALS in enhancing P-wave signal strength and AV ratios, indicating its potential as a valuable tool for enhancing clinical screening and diagnosis of atrial arrhythmias. Additionally, the results from this research may induce diagnostic specialists to incorporate the optimized lead configuration of ALS into long-term, ambulatory, telemetric, and mHealth devices to improve monitoring and ensure accurate diagnosis of various atrial arrhythmias. This approach is reliable for precisely diagnosing atrial arrhythmias in real-time clinical settings.