Artificial Intelligence With DEep Learning on COROnary Microvascular Disease
Artificial Intelligence With DEep Learning on COROnary Microvascular Disease
Despite the progress made in the management of myocardial infarction (MI), the associated morbidity and mortality remains high. Numerous scientific data show that damage of the coronary microcirculation (CM) during a STEMI remains a problem because the techniques for measuring it are still imperfect. We have simple methods for estimating the damage to the MC during the initial coronary angiography, the best known being the calculation of the myocardial blush grade (MBG), but which is semi-quantitative and therefore not very precise, or more precise imaging techniques, such as cardiac MRI, which are performed 48 hours after the infarction and which make the development of early applicable therapeutics not very propitious. Finally, lately, the use of special coronary guides to measure a precise CM index remains non-optimal because it prolongs the procedure. However, the information is in the picture and this information could allow the development of therapeutic strategies adapted to the patient's CM. Indeed, the arrival of iodine in CM increases the density of the pixels of the image, this has been demonstrated by the implementation in 2009 of a software allowing the calculation of the MBG assisted by computer. But the performances of this software did not allow its wide diffusion. Today, the field of medical image analysis presents dazzling progress thanks to artificial intelligence (AI). Deep Learning, a sub-category of Machine Learning, is probably the most powerful form of AI for automated image analysis today. Made up of a network of artificial neurons, it allows, using a very large number of known examples, to extract the most relevant characteristics of the image to solve a given problem. Thus, it uses thousands of pieces of information, sometimes imperceptible to the naked eye. We hypothesize that a supervised Deep Learning algorithm trained with a set of relevant data, will be able to identify a patient with a pejorative prognosis, probably related to a microcirculatory impairment visible in the image.
The aim of this study is to develop an algorithm capable of identifying patients with poor prognosis criteria at the time of hospitalization for STEMI, despite successful revascularization by analyzing coronary angiography images using supervised Deep Learning type artificial intelligence methods.
The protocol will be subdivided into 4 steps:
- Step 1: Patient selection
Data mining to identify and select patients via PMSI data. Patients will be contacted by telephone follow-up to check the participation agreement and collect the primary outcome. Other data from the patient's medical file will be collected through PREDIMED.
- Step 2: Data annotation
To identify for each patient with successful revascularization according to the usual criteria (TIMI Flow = 3, MBG = 2 or 3 and ST segment resolution > 70%) whether or not he or she presents, at the time of hospitalization for STEMI, pejorative evolution criteria defined by the occurrence of death or rehospitalization for heart Failure at the time of follow-up . This step requires the expertise of an angioplastician and will result in the generation of a database of 600 cases. To train the algorithm to recognize images in the context of STEMI revascularization, 1000 normal coronary angiographies performed in a stable disease context will also be identified.
- Step 3: Development of a new method for analyzing coronary angiography images to identify patients with non-optimal revascularization.
Develop using Tensorflow/Keras libraries a supervised Deep Learning AI algorithm trained to identify patients with non-optimal revascularization (patient with poor prognosis). The algorithm will be based on convolutional neural network methodology and the model will be trained using data from the two previous steps. All or part of the sequence of interest will be used at the input of the model which will propose at the output a probability of good or bad prognosis of the patient.The 1000 complementary coronary angiographies will be used to artificially increase the learning base by increasing the number of cases or will be exploited for a transfer learning method.
- Step 4: Evaluation of the pathophysiological hypothesis.
The main weakness of AI is the "Black Box". That is, the algorithm can predict correctly without knowing how. It is then difficult to link the result to a physiopathological phenomenon and to develop therapeutics. Here we will evaluate the correlation of the algorithm's result with the reference method for measuring CD used in the patients of the Guardiancory study (NCT03087175).
Inclusion Criteria:
Exclusion Criteria:
gbarone@chu-grenoble.fr+33476765172
pperetout@chu-grenoble.fr+33476766700