(G-259) Novel Deep Learning Frameworks for Real-time R-Peaks Detection in the Corrupted ECG Signals, Denoising of Heart Sounds and Smartwatch-based Covid-19 Detection

Introduction::
Note: This abstract best fits the Specialty track under "Deep Learning Strategies"

Abstract:
The current work presents three latest studies, models and results from our group related to deep learning in medical applications. (i) A self-attention MHDNet: a novel deep learning model for the detection of R-peaks in the electrocardiogram signals corrupted with magnetohydrodynamic effect, (ii) an end-to-end deep learning framework for real-time denoising of heart sounds for effective computer-aided auscultation, and (iii) PCovNet+: A CNN-VAE anomaly detection framework with LSTM embeddings for smartwatch-based COVID-19 detection.

Introduction:
Artificial Intelligence (AI) has revolutionized the fields of computer vision, Natural Language Processing (NLP), and robotics, and it holds enormous potential for revolutionizing the future of medicine, healthcare and biomedical research. Particularly, the field of healthcare is currently experiencing a paradigm shift with the integration of deep learning techniques in diagnostics, treatments, and drug discoveries. In this context, our research group along with collaborators has made significant advancements in development of new deep learning and AI models for diverse medical applications. In this work, we present several examples of latest work in our group on applications of deep learning models in medical applications.

Materials and Methods::
A self-attention MHDNet model [1] was developed for the detection of R peaks in electrocardiogram (ECG) signals distorted by the magnetic fields induced by MRI scans. Here, a self-attention U-Net based model was employed, leveraging the advantages of the architecture to accurately identify the R peaks despite the interference caused by the MRI magnetic field. In the second study, a U-Net based model, was developed effectively mitigating noise and enhancing the quality of heart sound recordings for effective computer-aided auscultation in Unseen Noise. In the third study, we introduced PCovNet+, a deep learning framework tailored for smartwatches and fitness trackers, enabling the detection of COVID-positive patients through the monitoring of Resting Heart Rate (RHR) anomalies. Our framework employed a convolutional neural network (CNN)-based variational autoencoder (VAE) architecture, augmented by a long short-term memory (LSTM) network to generate latent space embeddings.

Results, Conclusions, and Discussions::
Results and Discussions
The developed Self-Attention MHDNet model achieved a recall and precision of 99.83% and 99.68%, respectively, for the ECG data acquired in a 3T setting, while 99.87% and 99.78%, respectively, in a 7T setting. The U-Net based proposed framework effectively suppressed background noise in both unseen real-world data and synthetically generated noisy heart sound recordings, improving the signal-to-noise ratio (SNR) level by 5.575 dB on an average using only 1.32 M parameters. It outperformed the current state-of-the-art U-Net model with an average SNR improvement of 5.613 dB and 5.537 dB in the presence of lung sound and unseen hospital noise, respectively. The PCovNet+ framework detected COVID-19 infection for 74% of the subjects (47% presymptomatic and 27% post-symptomatic detection).

Conclusions
Collectively, these research endeavors exemplify the promising prospects of deep learning in healthcare including disease detection, monitoring, and diagnosis, offering valuable insights and potential solutions for the benefit of patients and healthcare practitioners alike. the PCovNet+ framework successfully detected COVID-19 infection for 74% of the subjects (47% presymptomatic and 27% post-symptomatic detection)

Acknowledgements: This project was partly funded by the Qatar National Research Fund under Grant NPRP12S-0310-190276. The statements made herein are solely the responsibility of the authors. The authors also acknowledge the contributions of the team members Shams Nafisa Ali, Samiul Based Shuvo, Mozjjem Hossain Chowdhury, Farhan Fuad Abir and other student members in the group.