A new AI accurately and quickly predicts the risk of heart attack over 5 years


You may also like this partner content (after the ad)

A team at Cedars-Sinai Medical Center in Los Angeles has developed a new artificial intelligence algorithm capable of predicting the risk of suffering a heart attack in the next five years. To do this, it relies on a precise measurement of the deposits of fatty plaques which accumulate on the walls of the arteries and which lead to atherosclerosis. The verdict falls in just a few seconds.

To predict the probability When a myocardial infarction (more commonly known as a heart attack) has occurred, doctors usually do a computed tomography (CTA) angiogram. This is a 3D imaging examination that allows you to observe the vascular network and detect any abnormalities; it thus makes it possible to estimate the quantity of fatty deposits which have accumulated in the coronary arteries. This buildup of coronary plaque can lead to narrowing of the arteries, or even total blockage, which prevents blood from reaching the heart, increasing the likelihood of suffering a heart attack.

However, this measurement of plaque volume is not systematic, as Professor Damini Dey of the Cedars-Sinai Biomedical Imaging Research Institute explains: “there is no fully automated way to do this […], it takes at least 25 at 65 minutes to an expert”. In contrast, the algorithm developed by Dey and his colleagues only takes five to six seconds to quantify coronary plaque from CTA images. This study constitutes the first validation of a deep learning approach for the quantification of atherosclerosis from coronary angiography images.

Efficiency similar to that of human diagnosis

To train their algorithm, the researchers formed a set of six cohorts (921 patients in total), from five different countries, all of whom underwent CT coronary angiography between 2010 and 2019 (via several different CT scanners and protocols). The images had previously been reviewed by expert human readers. The clinical spectrum of coronary artery disease ranged from stable angina to acute myocardial infarction, up to post-myocardial infarction convalescent stage. The severity of the stenosis was scored from 0 to 5 depending on the level of narrowing of the artery.

The algorithm is based on a new network neural network based on ConvLSTM — a hierarchical convolutional long-term memory network — which is the first to accurately quantify coronary stenosis and the volumes of all components of atherosclerotic plaque. It works by first highlighting the coronary arteries in 3D images and then identifying blood and plaque deposits in the coronary arteries.

correlation estimates plate volume IA experts

Graphiques de corrélation (A) et de Bland-Altman (B) comparant le volume total de plaque (en mm³) mesuré par apprentissage en profondeur par rapport à des lecteurs experts dans le cas de 1901 lésions examinées. © A. Lin et al.

” data-image-title=”correlation-estimations-IA-experts” data-large-file=”https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts.jpg” data-lazy-sizes=”(max-width: 800px) 100vw, 800px” data-lazy-src=”https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts.jpg” data-lazy-srcset=”https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts.jpg 900w, https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts-300×121.jpg 300w, https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts-500×201.jpg 500w, https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts-472×190.jpg 472w, https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts-390×157.jpg 390w” data-medium-file=”https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts-300×121.jpg” data-orig-file=”https://trustmyscience.com/wp-content/uploads/2022/03/correlation-estimations-IA-experts.jpg” data-orig-size=”900,362″ data-permalink=”https://trustmyscience.com/nouvel-algorithme-predit-precisement-rapidement-risque-infarctus/hires-15/” height=”322″ src=”data:image/svg+xml,%3Csvg%20xmlns=’http://www.w3.org/2000/svg’%20viewBox=’0%200%20800%20322’%3E%3C/svg%3E” width=”800″>

Correlation (A) and Bland-Altman (B) plots comparing total plaque volume (in mm³) measured by deep learning against expert readers in the case of 1901 lesions examined. © A. Lin et al.

The effectiveness of the tool has been validated on a set of test images of 238 patients. The results are unequivocal: the researchers report “excellent agreement between deep learning and expert readers” for the estimation of total plaque, calcified plaque and non-calcified plaque volumes. The concordance was also good for the evaluation of the diameter of the stenosis.

All for a record analysis time: “ The average deep learning plate analysis time per patient was 5,65s when the calculation was performed using a graphics processing unit and 3,82 min with the use of a central processing unit ”, specify the researchers in The Lancet Digital Health. For comparison, the average analysis time by the experts was 30,66 min per patient.

A reliable and non-invasive prediction tool

The team specifies that the algorithm has been shown to be equally effective , when its results were compared with those obtained from images taken during two invasive tests considered to be very accurate for evaluating plaque and narrowing of the coronary arteries: intravascular ultrasound and catheter coronary angiography. “The deep learning-based analysis showed excellent agreement and correlation with intravascular ultrasound for measurements of total plaque volume and the minimum luminal zone”, underline the authors of the study.

t/uploads/2022/03/risk-infarction-volume-plaque-300×181.jpg” data-orig-file=”https://trustmyscience.com/wp-content/uploads/2022//risk-infarction-volume-plaque.jpg” data-orig-size=”900,514″ data-permalink=”https://trustmyscience.com/new-algorithm-predicts-precisely-rapidly-risk-infarction/hires-17/”>

plaque volume infarction risk
Myocardial infarction prognosis from total plaque volume estimated by deep learning . © A. Lin et al.

Once the measurement capabilities of the tool had been assessed, the researchers wanted to test its prognostic capabilities. The algorithm had to estimate the risk of a fatal or non-fatal myocardial infarction occurring within five years, based on scans of a sample of 1611 patients with stable chest pain (including 41 actually suffered a heart attack in the years following their examination). Again, the program showed great accuracy. It appears from the study that patients with a total plaque volume of 238,5 mm³ or more had a 7 times higher risk of having a myocardial infarction than patients with a lower volume.

Further research will be needed to train the algorithm from a larger and more varied population of patients. But this first study is a new example of the incredible performance of artificial intelligence in image analysis, which will most certainly be put to use in the future of medicine. “ Further studies are needed, but we may be able to predict whether a person is likely to have a heart attack and in what timeframe, based on the amount and composition of plaque imaged by this standard test,” Dey said.

Source: A. Lin et al ., The Lancet Digital Health