Recognizing the causes of LVH: an AI does as well as cardiologists

Los Angeles, USA – When echocardiography reveals left ventricular hypertrophy, the physician is faced with a tedious and time-consuming task: determining its extent and its cause. Artificial intelligence (AI) can make it easier, as demonstrated by a US study that not only accurately measured the geometric characteristics of the ventricle but also identified underlying conditions, including aortic stenosis, hypertrophic cardiomyopathy and cardiac amyloidosis. The work has been published in JAMA Network[1].

Highlight of this study by the team of Grant Duffy (Cedars-Sinai Medical Center in Los Angeles): A large set of data from medical videos had never before been published. “With nearly 24,000 cardiac ultrasound videos, this is an impressive publication, if only for the enormous amount of data used by the authors – such as training, validating and testing their algorithms – in the different stages of their work”, confirms Dr Jackie Mathe leader of the Applied Machine Learning group at the Fraunhofer Heinrich-Hertz Institute in Berlin interviewed by Medscape

AI saves doctors a lot of time and effort

If Jackie Ma sees the importance of AI, the Pr David Ouyang, last author of the study, gives it back its place: “It is important to specify that the algorithms leave the decision to the doctors. They simply offer them help by suggesting such and such a diagnosis,” one can read on the tctMD cardiology website.

Internationally, Germany has a fairly strong position in the field of AI research, development and applications, reports Jackie Ma. Grant Duffy, but rather to analyze ECGs. He has also developed algorithms for the analysis of proteins and the genome, and that of EEGs and cerebral MRIs, chest X-rays and the diagnosis of certain tumours. These researchers were also interested in the Covid-19 pandemic, through the research of the personal risk of contamination according to the distance with an infected person and the time spent near him.

A critical point: the acquisition of reference data

“We work closely with different clinics, but strict data protection makes it difficult to access appropriate source material,” says Jackie Ma. The study authors share the same experience: One of the main challenges of using AI in healthcare is the lack of benchmarks. »

However, the authors were fortunate to be able to source data from their own clinic as well as from a Stanford center, one of whose specialties is cardiac hypertrophy. They retained on the one hand videos of the long parasternal axis, on which one can see the two ventricles, the root of the aorta and the left atrium, and on the other hand, images showing the 4 cavities from the apex.

Echocardiography, for a diagnosis of first choice

In accordance with the recommendations of learned societies, echocardiography is the method most used to diagnose hypertrophy, explain the authors of the study. Jackie Ma underlines the related problems: “Due to the large amount of information they deliver, echocardiograms are difficult for AI users to manage: they require a large investment in terms of time, capacity computing and storage space. »

Grant Duffy and his collaborators used part of the videos to train their algorithm, indicating to it, after each passage, whether it had reached the right or the wrong result (see box). The analysis reports and the remarks attached each time served as a point of reference. “The question is always: where do the analyzes and annotations come from? The ideal would be for as many doctors as possible to give their opinion,” says Jackie Ma.

This process would be particularly beneficial for patients with hypertrophy because even experts find it difficult to differentiate between different pathologies that modify the heart in morphologically similar ways. What is also confusing is that the symptoms vary from person to person and are sometimes milder, sometimes more severe. In addition, the measured values ​​vary due to irregularities in the filling time and in the heart rate.

A reliable examination is nevertheless important, because it determines the continuation of the procedure because of its prognostic importance. It is thus possible to assess the risk of sudden cardiac death and to determine which patients require a defibrillator. Additionally, classification based on genetic differences would also be possible.

An algorithm that enhances the diagnostic potential of ultrasound

“We therefore sought to find out if it was possible to bring out the hidden potential of echocardiography by combining it with a variant of artificial intelligence, deep learning (see box)”, explain Grant Duffy and his team for justify the interest of their retrospective study.

In a prospective comparison with two trained cardiologists, the algorithm showed even slightly better results.

The approach was successful: the algorithm spotted deviations, even when they were subtle. For intraventricular wall thickness, the mean error was only 1.2 mm. It was 2.4 mm for the diameter of the left ventricle and 1.4 mm for the thickness of the posterior wall. In a prospective comparison with two trained cardiologists, the algorithm showed even slightly better results. It performed equally well when capturing data from other US clinics and other countries. According to the researchers, this demonstrates that its relevance continues across continents and different health systems.

A process that mimics the visual cortex

Investigators searched for the primary condition using a special deep learning method that draws on the visual cortex, the Convolutional Neural Network. This “convolutional neural network” was thus able to identify with a high level of certainty cardiac amyloidosis, hypertrophic cardiomyopathy and aortic stenosis separately from other possible causes.

What these underlying conditions have in common is that they generate a chronic overload that the heart tries to overcome by remodeling.

In many patients, the disorder is of systemic origin: the heart muscle has to work against increased pressure, for example in the case of aortic stenosis, but especially in the case of high blood pressure, which initiates remodeling in 60% of patients. .

Triggering genetic defects in muscle fibrils

But the thickening can also stem from pathological processes within the heart itself, as in hypertrophic cardiomyopathy. The latter finds its origin in hereditary mutations: more than 1500 are known in genes coding mainly for proteins of the sarcomere.

Diagnosticians must also distinguish another series of diseases that directly weaken heart tissue: amyloidosis. Only a few of them develop due to inherited genetic defects. Most are due to diseases of the bone marrow or lymph glands, or chronic inflammation such as in rheumatoid arthritis, Crohn’s disease or ulcerative colitis. As a result, misfolded protein fibrils are deposited between the cells. About thirty proteins have been discovered to date as being the cause of such amyloidosis.

Finally, if their presence is rare, other generally congenital disorders are not excluded, such as Fabry’s disease, Friedreich’s ataxia or MELAS syndrome.

A rare disease, but probably more common than you think

Doctors can easily miss the primary condition because it presents as a consequence of “normal” high blood pressure or “ordinary” kidney disease, explains david ouyang on tctMD. “Cardiac amyloidosis probably deserves the paradoxical title of ‘common rare disease’, in that it is more widespread than the published figures suggest. »

Hypertrophic cardiomyopathy also presents a diagnostic challenge due to its lack of uniformity: it sometimes spreads diffusely, with walls up to 60 mm thick, but sometimes only circumscribed areas are slightly thickened.

The authors consider their model as a platform technology because it is suitable for screening for all kinds of diseases, such as valvulopathies or lesions due to chemotherapy. “It can be fine-tuned to find patients who would not have been diagnosed in routine practice,” adds David Ouyang.

the deep learning

Deep learning is a sub-domain of machine learning. While the latter’s algorithms run through mathematical decision trees, deep learning dives, so to speak, deeper. It is based on neural networks whose operation is inspired by the human brain. These networks consist of an arrangement of input and output neurons, linked by a variable number of intermediate layers.

Algorithms thus develop their ability to independently determine the criteria for correct identification, by constantly integrating new content and taking other paths. Unlike machine learning, the programmer no longer intervenes in these processes: he simply feeds them with the basic information, from which the algorithms make predictions and make decisions. In practice, they are suitable for finding patterns, for example for image analysis but also, beyond medicine, for the recognition of faces, objects or speech.

The calculations are fully automated and lead to precise and reproducible measurements.

This article was originally published on and titled „Beindruckend“ – künstliche Intelligenz macht Kardiologen Konkurrenz: Sie erkennt Ursachen einer Hypertrophy mindestens so gut . Translation/adaptation by Dr. Claude Leroy.

Follow Medscape in French on Twitter.

Follow | Medscape Cardiology on Twitter.

of the

We would like to give thanks to the author of this short article for this awesome web content

Recognizing the causes of LVH: an AI does as well as cardiologists

Find here our social media accounts as well as other pages related to it.