New AI Models Diagnose Heart Conditions Without Cardiologists, Potentially Helping Underserved Areas

Artificial intelligence (AI) has its roots in the 20th century and has burst into full bloom in the 21st. There are not many places where this trend is more visible than in the field of medicine. AI can and has been applied in all aspects of healthcare including prevention, screening, diagnosis, treatment, and prognostication.Heart

Cardiology as a field uses a large amount of biosignals which need to be continually and regularly monitored and AI can offer these functions and services.

Initially, AI may be used to interpret data, however, it is possible that it can also be used to provide specialist care in areas that have scarce medical resources and specialists. This is a more challenging venture but it may accelerate the growth of the healthcare industry.

Read Also: Artificial Intelligence Boosts Breast Tumor Detection Rates in Mammography Screening: Findings from a Swedish Study

Background

According to the US Centers for Disease Control (CDC), about 46 million Americans have limited access to specialist services. The World Health Organisation (WHO) also estimated that 3.5 billion people have limited access to specialist services. Chest pain is recorded to be the number one cause of death globally.

In resource-limited settings, healthcare personnel have to make a cardiovascular diagnosis with what they garner from chest auscultation, point-of-care ultrasound devices (POCUS), and limited echocardiogram images. Because all these modalities require specialist attention, they’re not able to make a diagnosis as quickly as possible to aid immediate intervention.

Three deep learning models

A study has discovered 3 well-performing deep learning AI models that accurately identified aortic stenosis, determined the presence of heart failure, and classified murmurs. Initially, there were 26 models which were tested but only 3 were able to perform better than cardiologists and prior AI models. All clinical information was obtained from readily available 2-D Echocardiogram images and videos and POCUS and all models went through training and testing.

They developed 15 models to assess the severity of aortic stenosis from 2-D echocardiogram images. The analysis included convoluted neural networks (CNN) which ensured that they were able to identify the model with the highest accuracy. The model, Efficient Net, was the most accurate.

Amongst the 6 deep learning models built for the identification of heart failure via ejection fraction, the EchoSwin base model was the most accurate. It functioned by identifying changes in left ventricular volume. Data were analyzed from 2-D echocardiogram videos.

Read Also: Prognosticating Lung Cancer with Artificial Intelligence Is Now Possible

Furthermore, they built 5 models to identify and classify cardiac murmurs. The aim was to build a cost-effective digital stethoscope that converts, amplifies, filters, and digitally processes heart sounds which was projected for the identification of murmurs on a device called “Virtual Cardiologist”. The digital stethoscope was trained to identify different heart conditions; Murmur Absent, Aortic Stenosis, Pulmonary Stenosis, Tricuspid Stenosis, Ventricular Septal Defect, Mitral Regurgitation, Pulmonary Insufficiency, Tricuspid Insufficiency, and Still’s Murmur. The sounds were processed through LSTM Recurrent Autoencoder which was the most accurate of them all.

Finally, a portable handheld device containing a mobile application called “The Virtual Cardiologist” and a computer application were used to store the AI models that performed better and they’re able to receive and analyze information from either a POCUS or digital stethoscope.

Clinical significance

The benefits of these deep learning models are that healthcare personnel can have a handy device that is accurate and inexpensive and that has the ability to identify various cardiac conditions in resource-limited settings. This allows for the comprehensive cardiac assessment of patients in the absence of specialists.

However, additional research will require deeper insights into the AI models to improve efficiency.

Clinical trials that include patients with varying presentations, age groups, and races are still needed. Also, studies are needed to formulate deep learning models that can go on further to identify other complex heart conditions like valvular diseases and cardiomyopathies.

Read Also: Artificial Intelligence Accurately Diagnoses Brain Tumors in Austrian Study

References

Mehrotra-Varma, S. (2024). The Virtual Cardiologist: Three Deep Learning Pipelines in an Inexpensive Portable Device and Web/Mobile Application for Rapid Cardiovascular Diagnosis and Clinical Decision-Making. [Preprint]. https://doi.org/10.1101/2024.05.27.24307981

FEEDBACK:

Want to live your best life?

Get the Gilmore Health Weekly newsletter for health tips, wellness updates and more.

By clicking "Subscribe," I agree to the Gilmore Health and . I also agree to receive emails from Gilmore Health and I understand that I may opt out of Gilmore Health subscriptions at any time.