Artificial selection

Artificial intelligence tool outperforms humans in detecting heart disease

Doctors at the Smidt Heart Institute in the US have developed an artificial intelligence (AI) algorithm that detects life-threatening heart disease more accurately than trained cardiologists.

Image Credit: Blue Planet Studio

The heart is a dynamic organism that reacts in various ways to physiological stress. Increases in left ventricular heart muscle thickness (hypertrophy) resulting from age are difficult to separate from diagnosed conditions such as hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis using routine imaging techniques .

However, deep learning algorithms in medical imaging applications promise to identify clinical phenotypes with greater accuracy than human experts.

Our AI algorithm can identify patterns of disease that cannot be seen with the naked eye, then use those patterns to predict the correct diagnosis… The algorithm identified high-risk patients more accurately than the well-trained eye of a clinical expert.

Dr. David Ouyang, Cardiologist, Smidt Heart Institute

A Brief Introduction to Deep Learning Algorithms

Deep learning is a type of AI learning method based on neural networks that mimic the way humans learn.

Neural networks work by accepting an input of data and letting the model construct an interpretation from that data. Deep learning algorithms are structured in a hierarchy of increasing complexity that differentiates them from conventional linear machine learning models.

Humans learn by recognizing features in their environment and representing them as abstract patterns in their own minds. The more humans learn, the more layers of abstraction they add to their mental models. So, for example, plants become abstract into trees, which are themselves abstract into oaks, and so on.

Similarly, each algorithm in the hierarchy of a deep learning model applies a transformation to the data it receives to produce an output as a statistical model. The process continues through successive layers until a certain degree of precision is achieved. The higher the number of layers processed by the model, the “deeper” the algorithm is.

The main advantage of deep learning is that the model builds its feature sets without supervision, which generally leads to faster and more accurate predictions.

This is because deep learning algorithms create complex statistical models from their own output. Thus, they are powerful tools for processing large amounts of unstructured data. This is crucial, because most of the data produced today is unstructured (or unlabeled). Images produced by cameras and medical equipment are an example of unstructured data.

Deep Learning Algorithms for Cardiovascular Diseases

Hypertrophic cardiomyopathy is a condition that causes thickening of the heart muscle, leading to arrhythmia and damage to the valves of the heart. Cardiac amyloidosis is a separate condition caused by amyloid deposits in heart tissue.

However, the two conditions often appear very similar on medical imaging using an echocardiogram. To further complicate diagnosis, in the early stages of development these diseases often resemble a healthy heart that simply changes shape and size with aging.

To address this challenge, the team led by Dr. David Ouyang of the Smidt Heart Institute at Cedars-Sinai Medical Center in Los Angeles developed a deep learning algorithm that quantifies and labels the heart’s left ventricular wall thickness. and predicts heart disease.

One of the most important aspects of this AI technology is not only the ability to distinguish abnormal from normal, but also to distinguish these abnormal conditions, because the treatment and management of each heart disease is very different. .

Dr. David Ouyang, Cardiologist, Smidt Heart Institute

Using long-axis parasternal echocardiogram videos, they performed frame-by-frame segmentation of left ventricular wall thickness. They then performed a beat-by-beat assessment of ventricular hypertrophy.

They used a three-dimensional convolutional neural network to predict the causes of left ventricular hypertrophy (LVH). They were also able to predict aortic stenosis (narrowing of the aortic valves) and cardiac amyloidosis.

Patients were selected from studies conducted at the Stanford Amyloid Center, Cedars-Sinai Medical Center (CSMC), Stanford Center for Inherited Cardiovascular Disease, CMSC Hypertrophic Cardiomyopathy Clinic for Hypertrophic Cardiomyopathy, and Advanced Heart Disease Clinic for Cardiac Amyloidosis between 2008 and 2020.

The patient pool included 23,745 patients.

Manual annotation by clinicians of intraventricular septum (IVS), LV internal dimension (LVID), and LV posterior wall (LVPW) measurements were used as training labels for the deep learning model.

The deep learning algorithm was trained on a dataset of 17,802 echocardiogram videos provided by Stanford Health Care.

The model design was done in Python, using the PyTorch deep learning library. A suitable DeepLabv326 architecture trained on long-axis parasternal images was used to distinguish important points used to measure ventricular dimensions.

The binary classification of video-based deep learning classifiers in the Smidt Heart Institute algorithm was similar in performance to a multi-brand, multi-class deep learning model for disease detection, but had the flexibility to be able to identify patients who had diagnoses of cardiac and aortic amyloidosis. stenosis.

The results from Cedars-Sinai Medical Center provide an opportunity for automated detection of cardiac structures in echocardiogram videos through deep learning. This opens up possibilities for the early detection and treatment of cardiovascular diseases.

References and further reading

Duffy, G. et al.(2022) High-Throughput Precision Phenotyping of Left Ventricular Hypertrophy with Cardiovascular Deep Learning. JAMA Cardiology [online] Available at:

Cedars of Sinai (2022) New artificial intelligence tool detects often overlooked heart disease. [online] Available at:

Disclaimer: The views expressed herein are those of the author expressed privately and do not necessarily represent the views of Limited T/A AZoNetwork, the owner and operator of this website. This disclaimer forms part of the terms of use of this website.