Changes in coronary plaque size can be predicted by using a deep learning model, a new study finds, which could help clinicians better understand the natural history of subclinical (non-obstructive) atherosclerosis. In findings presented at Cardiovascular Research Technologies (CRT) 2024 in Washington, D.C., the research team revealed results from its intravascular ultrasound (IVUS) study that administered statins as external validation to the study results. Presenting at the session, Hector M. Garcia-Garcia MD, PhD, of Georgetown University, showed that the models predicted size changes correctly (>80%) of the length of the pullback in >75% patients. There were no variables (patient, laboratory or allocation arm) associated with the prediction of the algorithm, Dr Garcia-Garcia added during Saturday’s Late Breaking Clinical Science session. High predictive accuracy “The best parameter for predicting plaque size changes, either progression or regression, was the plaque burden derivative,” he said. “Using this parameter, deep learning models showed high accuracy for the prediction of the progression/regression of the coronary artery disease in nonculprit vessels of patients with acute myocardial infarction.” Garcia-Garcia, also the Scientific Lead of Invasive Imaging at the MedStar Cardiovascular Research Network, explained that substantial residual risk of plaque progression could be seen after combination treatment with statin and the proprotein convertase subtilisin/kexin type 9 inhibitor (PCSK9i). Referencing one study, in which approximately 20% of patients demonstrated atheroma progression, despite achieving low-density lipoprotein (LDL) cholesterol ≤70 mg/dL, Garcia-Garcia pointed to a direct relationship between the burden of coronary atherosclerosis, its progression, and adverse cardiovascular events. Methodology Patient data were taken from the Integrated Biomarkers and Imaging Study-4 (IBIS-4) and the Effects of the PCSK9 Antibody Alirocumab on Coronary Atherosclerosis in Patients With Acute Myocardial Infarction (PACMAN-AMI) study. Baseline characteristics for patients from the IBIS-4 study (n=103) included a mean age of 58.2 ±10. years) of which 9.7% were female. Baseline characteristics for patients from the PACMAN-AMI study (n=265) included a mean (SD) age of 57.8 (9.3) of which 16.2% were female. These patients, who had experienced an acute myocardial infarction (AMI) underwent successful primary percutaneous coronary intervention (PPCI) and who had two non-infarct-related arteries with angiographic evidence of atherosclerosis (20% to 50% diameter stenosis). Deep learning model development In creating the deep learning model, the first derivative (∂f) of the plaque burden was computed using a central finite difference approximation. Frame-wise features that were computed were the lumen area, vessel area (VA), plaque area (PA) and plaque burden defined as the percentage of the VA occupied by PA. Additional model training included the IVUS images of 103 acute ST-elevation myocardial infarction patients, who were asked to take 20 mg of rosuvastatin from baseline over 2 weeks followed by 40 mg of rosuvastatin over 13 months. External validation of the deep learning model’s capabilities used IVUS images from 300 patients, who were either taking no statins and had an LDL-Clevel of >125 mg/dL (>3.2 mmol/L) or were taking statins and had a cholesterol level of LDL-C >70 mg/dL (>1.8 mmol/L) at baseline. Less than 24 hours after undergoing PCI, these patients were then either put on a course of alirocumab (150 mg for 2 weeks) and 20 mg of rosuvastatin or took the placebo for 2 weeks and then 20 mg of rosuvastatin of which IVUS images were collected after 52 weeks. The primary endpoint was defined as the change in percent atheroma volume (PAV) by greyscale IVUS. The models’ high accuracy was not associated with any variables (patient, laboratory or allocation arm) associated with the prediction of the algorithm, Garcia-Garcia said. Photo Credit: Bailey Salimes/CRTonline.org Photo Caption: Hector M. Garcia-Garcia MD, PhD, discusses the results of an intravascular ultrasound-based deep learning model Saturday at CRT 2024 in Washington, D.C.