New research suggests that a machine learning model that combines cardiac magnetic resonance imaging (MRI) features and clinical data may offer enhanced risk stratification beyond traditional models for patients with ST-segment elevation myocardial infarction (STEMI).
For the retrospective study, recently published in Radiology, researchers developed and compared the machine learning model with a clinical model, the Global Registry of Acute Coronary Events (GRACE) score and the Thrombolysis in Myocardial Infraction (TIMI) score. The cohort of 1,066 patients with STEMI (mean age of 58.15) included 384 patients in the external validation test set, according to the study.
For the prediction of major adverse cardiovascular events (MACE), the study authors found that the machine learning model demonstrated a 91 percent AUC in comparison to 86 percent for the clinical model, 66 percent for the GRACE model and 62 percent for the TIMI scoring system.
The machine learning model provided higher sensitivity than the clinical model for MACE prediction (82.7 percent vs. 80.3 percent), and greater than 30 percent higher sensitivity in comparison to the TIMI (51.8 percent) and GRACE scoring systems (42 percent), according to the researchers. The researchers also pointed out higher specificity with the machine learning model (84.5 percent) in contrast to the clinical model (74.9 percent) and TIMI scoring system (68.3 percent).
“ … We successfully developed and externally tested a machine learning (ML) model that integrates comprehensive cardiac MRI and clinical parameters for the prediction of long-term major adverse cardiovascular events in patients with ST-segment elevation myocardial infarction. The ML model demonstrated excellent predictive performance compared with existing clinical risk scores and traditional models, showing strong potential for personalized risk assessment,” noted lead study author WeiHui Xie, M.D., Ph.D., who is affiliated with the Department of Radiology at the Renji Hospital and the School of Medicine at Shanghai Jiao Tong University in Shanghai, China, and colleagues.
Utilizing a scoring system to identify those at low risk (< 14), intermediate risk (14 < 27) and high risk ( > 27) for MACE, the study authors said findings with the machine learning model revealed more than double the MACE risk for those with intermediate risk and over a fivefold higher risk for MACE in the high-risk category.
“By quantifying individualized risk thresholds, this tool enables tailored clinical decision-making—for example, intensified monitoring for patients with high risk versus deescalation for individuals with low risk—to improve patient care and optimize resource allocation,” added Xie and colleagues.
Three Key Takeaways
• Improved MACE prediction beyond traditional scores. A machine learning model integrating cardiac MRI features with clinical data achieved a 91 percent AUC for long-term MACE prediction in STEMI patients, outperforming a clinical model (86 percent), the GRACE score (66 percent), and the TIMI score (62 percent).
• Superior sensitivity and specificity. The ML model demonstrated higher sensitivity (82.7 percent) than the clinical model (80.3 percent) and markedly greater sensitivity than GRACE (42 percent) and TIMI (51.8 percent), along with higher specificity (84.5 percent vs. 74.9 percent for the clinical model and 68.3 percent for TIMI).
• Actionable risk stratification for personalized care. Risk score thresholds (<14 low, 14–27 intermediate, >27 high) identified more than a twofold increase in MACE risk for intermediate-risk patients and over a fivefold increase for high-risk patients, supporting tailored monitoring intensity and resource allocation for patients with STEMI.
In an accompanying editorial, Jerome Garot, M.D., Ph.D., and Suzanne Duhamel, M.D., noted the promise of the machine learning model for its externally validated combination of MRI features and clinical data.
“This risk score could be readily implemented in clinical practice as a bedside risk assessment tool to predict individual patient risk after MI. As such, the so-called black box effect seems limited, and this technique has the potential to mitigate the understandable reluctance of practitioners and gain clinical acceptance,” posited Drs. Garot and Duhamel, who are affiliated with the Department of Cardiovascular Imaging at the Cardiovascular Institute Paris Sud in Massy, France.
(Editor’s note: For related content, see “What New CCTA Research Reveals About Quantitative Plaque Assessment and Predicting MACE,” “Is MRI Assessment of Right Ventricular Global Longitudinal Strain Prognostic for Heart Failure?” and “FDA Clears Updated AI Software for Generating Inline Mapping with Cardiac MRI.”)
In regard to imitations, the study authors acknowledged that the Chinese cohort may thwart extrapolation of the study findings to broader populations. They also conceded that a lack of availability for routine cardiac MRI may hamper use of the machine learning model in certain settings.
