AI-automated MVO radiomic analysis effectively predicted MACE risk and outperformed conventional quantitative assessment in patients with STEMI.
Key Findings
Methods
An AI-based radiomic model for microvascular obstruction (MVO) analysis was developed and validated across multiple centers in 843 STEMI patients.
Multicenter retrospective study spanning June 2013 to December 2023
843 patients with STEMI and MVO who underwent cardiac MRI were included (median age 60 years [IQR, 51-67 years]; 760 male patients)
Dataset was split into training set (n = 387), validation set (n = 166), and external test set (n = 290)
A previously developed AI model was applied for automated MVO segmentation, followed by radiomic feature extraction
Results
From 1595 extracted radiomic features, LASSO regression identified six features used to construct a radiomic score (radscore).
1595 radiomic features were extracted from AI-segmented MRI scans
Least absolute shrinkage and selection operator (LASSO) regression was used for dimensionality reduction
Six features were selected for constructing the final radscore
The radscore was designed to capture heterogeneity of microvascular injury beyond simple volume quantification
Results
Patients who experienced MACEs had significantly higher radscores than those who did not.
190 of 843 patients experienced MACEs (including cardiovascular death, myocardial reinfarction, malignant arrhythmia, and hospitalization for heart failure)
Mean radscore for patients with MACEs was -0.98 ± 0.50 (SD) versus -1.42 ± 0.50 for those without MACEs
Difference was statistically significant (P < .001)
Results
The radscore was an independent predictor of MACEs with a hazard ratio of 4.20.
Hazard ratio for radscore predicting MACEs was 4.20 (95% CI: 3.19, 5.53; P < .001)
The radscore demonstrated greater prognostic value compared with conventional MVO volume quantification
Restricted cubic spline analysis was performed to examine potentially nonlinear relationships between radscore and MACE risk
Results
Integrating the radscore with conventional variables significantly improved prognostic performance over conventional variables alone.
C index for conventional variables alone was 0.77 (95% CI: 0.73, 0.81)
C index increased to 0.80 (95% CI: 0.77, 0.83) when the radscore was added to conventional variables
Improvement in C index was statistically significant (P < .001)
The radscore contributed to optimizing risk stratification beyond standard clinical assessment
What This Means
This research suggests that an artificial intelligence (AI) system can analyze cardiac MRI scans to better predict serious heart complications after a severe type of heart attack (STEMI). When someone has a STEMI, tiny blood vessels in the heart muscle can become blocked even after the main artery is reopened — a condition called microvascular obstruction (MVO). Traditionally, doctors measure how large this area of obstruction is, but this study found that analyzing the texture and complexity (heterogeneity) of the MVO region using AI and a technique called radiomics provides more useful information than simple size measurements alone.
The AI model was tested on 843 heart attack patients from multiple hospitals, and 190 of them went on to experience major adverse cardiovascular events (MACEs) such as heart failure hospitalization, dangerous heart rhythms, repeat heart attacks, or cardiovascular death. The AI-generated score (called a radscore) was significantly higher in patients who later had these complications, and it predicted outcomes better than conventional MVO volume measurements. When the radscore was combined with standard clinical information, the ability to predict who would experience complications improved from a C index (a measure of predictive accuracy) of 0.77 to 0.80.
This research suggests that capturing the complexity and spatial variation of microvascular damage — not just its size — provides important prognostic information. If validated further, this type of AI-assisted analysis could help clinicians identify heart attack patients who are at higher risk for future complications and may need more intensive monitoring or treatment, without requiring additional testing beyond what is already done with cardiac MRI.
Chen B, Li S, Xiang J, Wu C, An D, Tang L, et al.. (2026). AI-based Histologic Heterogeneity of Microvascular Obstruction at Cardiac MRI for Predicting MACEs: A Multicenter Study.. Radiology. https://doi.org/10.1148/radiol.252250