The old EKG readout may become a thing of the past. (AP Photo/Jeff Roberson)
A team of researchers has proposed a new automated cardiac screening system powered by artificial intelligence that could make cardiac diagnoses accessible to all without the need for a trained practitioner to read or interpret results.
The system, called CardioXNet, was described in a study published March 2 in IEEE Access, a journal of the Institute of Electrical and Electronics Engineers. Its design uses deep-learning techniques to detect five diagnoses automatically and is said to deliver near-perfect results.
Considering that not having manual preprocessing steps is key for any end-to-end mobile health care device for clinicians using deep-learning models, the proposed framework was also designed to not require any such steps.
The framework relied on recordings of all the sounds of the human heart, including its vessels and chambers, produced through a diagnostic technique called phonocardiography. Researchers intend to integrate CardioXnet in the future with digital stethoscopes or wearable devices with a cloud server connection for automatic cardiac signal classification using the team's pre-trained model in real time to aid clinicians in making their diagnostic decisions. Any potential barriers that could remain, such as regulatory clearance, are contingent on the final product's post-integration design.
Auscultation, the practice of listening to the sounds of the body, a process that is performed to examine certain systems, including the circulatory system and the respiratory system, is "The most efficient and cost-effective way to primarily diagnose cardiovascular diseases," said study author Mabrook S. Al-Rakhami, research chair of pervasive and mobile computing at King Saudi University in Riyadh, Saudi Arabia.
"However," he continued, "readily detecting cardiovascular diseases from auscultation is a pretty challenging task and requires years of practice."
For one, it requires excellent hearing and the medical ability to distinguish subtle differences in pitch and timing. Also, physicians can interpret the same auscultation differently, exacerbating this challenging task.
"The proposed model can be used as an auxiliary system by clinicians in the diagnosis and evaluation of cardiovascular diseases with high accuracy in health centers, especially in developing countries where other high-end diagnostic tests are insufficient," said Al-Rakhami, who is also a senior member of the IEEE.
The researchers reported that their framework can automatically predict cardiovascular disease using five auscultation classes: normal; aortic stenosis, or narrowing of the aortic valve opening; mitral stenosis, or narrowing of the mitral valve opening; mitral regurgitation, or leakage of the blood backward via the mitral valve; and mitral valve prolapse, where the two mitral valve flaps do not close as they should.
The study was limited, though, in the amount of phonocardiogram signals data, and actual patient data were not included within the dataset.
Instead, the researchers used raw phonocardiogram signals in audio files of heart sounds and murmurs, with a total of 1,000 recordings in the five classes collected from various sources including books and websites.
Monitoring heart conditions through phonocardiogram signals is becoming an increasingly widespread practice due to its simplicity and cost-effectiveness. It allows clinicians to more easily process recordings of all of the sounds that are made by the heart during a cardiac cycle, and extract salient features for identifying cardiac anomalies. Numerous studies have explored automated signal classification over the years, but, "Most of these have incorporated resource-intensive, time-consuming, complex signal processing-based feature extraction and heavy deep-learning networks," Al-Rakhami said.
The artificial intelligence-powered classification of valvular heart disease from phonocardiogram signals does not require manual extractions or preprocessing steps such as segmentation or augmentation. The team focused on the ability to extract the essential features directly from these signals, with minimal training parameters and electronic memory.
Cardiovascular disease constitutes a large portion of deaths globally. One person dies every 36 seconds in the U.S. from cardiovascular disease, according to the Centers for Disease Control and Prevention, and about 655,000 Americans die from heart disease each year, accounting for one in every four deaths.
Early diagnosis of cardiac conditions becomes increasingly crucial in the face of clinician shortages, as the struggle of health care professionals to meet growing demand from more patients has become widely acknowledged over the past several years, from the U.S. to developing countries such as Saudi Arabia, and rural areas are impacted even more.
The COVID-19 pandemic has further strained cardiovascular screening facilities in rural areas, Al-Rakhami noted. But CardioXNet could help close the primary care workforce gap globally in the critical area of cardiovascular disease, its creators say, including through faster diagnoses.
Features of CardioXNet also make it "more suitable for real-time wearable and mobile applications" than other networks, the researchers argue. Al-Rakhami highlighted some of the network's novel features: an end-to-end time of 54.60 milliseconds, a relatively low number of 0.67 million trainable parameters and 26 million floating-point operations per second with a minor memory space requirement of 7.96 megabytes.
"Despite having this surprisingly lower number of parameters, it performs on par with the state-of-the-art lightweight architectures like MobileNet, WaveNet, etc.," Al-Rakhami said. Cornell University researchers set MobileNet forth in 2018. Different teams have proposed at least two such architectures, including WaveNet, just last year.
The team touted CardioXNet's results, showing an average of 99.60% accuracy, 99.56% precision, 99.52% recall and 99.68% F1- score, a measure of a model's accuracy on a dataset, on both primary and secondary datasets.
Automation nixes the resource-intensive requirement of implementing deep learning-based frameworks, which hinders deployment in "low-resource point-of-care locations of the developing and under-developing countries," researchers said.
The framework combines deep learning architectures in a so-called convolutional recurrent neural network, which is said to generate better results, especially in audio signal processing. It involves two learning phases: representation and sequential residual learning, enabling automation directly from phonocardiogram recordings.
It extracts time-invariant features and converges extremely quickly in the representation learning phase — meaning more training wouldn't improve the model — and it emphasizes extracting temporal features in the sequential learning phase, according to Al-Rakhami.
The research team's future aim is to implement the proposed CardioXNet on multiple larger phonocardiogram datasets with diverse cardiovascular annotations, Al-Rakhami said.
"We would also like to integrate this architecture with a digital stethoscope to assist clinicians in their diagnostic decision in real-time," he added.
The study, "CardioXNet: A Novel Lightweight Deep Learning Framework for Cardiovascular Disease Classification Using Heart Sound Recordings," published March 2 in IEEE Access, was authored by Samiul Based Shuvo, Shams Nafisa Ali and Soham Irtiza Swapnil, Bangladesh University of Engineering and Technology; and Mabrook S. Al-Rakhami and Abdu Gumaei, King Saud University.