AI-Powered MRI Evaluations Predict STEMI Outcomes Better than Existing Risk Scores

Why It Matters

Improved risk prediction can personalize post‑STEMI care, potentially reducing mortality and rehospitalization. Demonstrating AI superiority over established scores signals a shift toward data‑rich clinical decision tools.

Key Takeaways

• ML model outperforms GRACE and TIMI scores
• Uses 67 variables, narrowed via feature elimination
• Tested on 1,066 Chinese STEMI patients, 40‑month follow‑up
• Achieved highest integrated AUC for MACE prediction
• Requires validation in diverse populations and other imaging

Pulse Analysis

Risk stratification after ST‑segment elevation myocardial infarction has long relied on scores such as GRACE and TIMI, which use a limited set of clinical variables. While useful, these tools often miss nuanced imaging information that can signal myocardial damage or recovery potential. The rapid adoption of artificial intelligence in cardiology offers a way to fuse high‑dimensional data—like cardiac MRI texture, ventricular volumes, and perfusion metrics—with traditional risk factors, creating a more granular patient profile.

In the recent Radiology publication, researchers from Renji Hospital built a machine‑learning pipeline that initially considered 67 variables, then applied recursive feature elimination to isolate the most predictive features. Trained on 682 patients and externally tested on 384, the model delivered an integrated area under the curve (AUC) that surpassed both GRACE and TIMI across all major outcomes, including cardiovascular death, repeat infarction, and heart‑failure rehospitalization. The median follow‑up of 40 months provided a robust temporal window, allowing the algorithm to demonstrate consistent discrimination throughout the study period.

The implications extend beyond a single study. If validated across diverse ethnic groups and with more accessible imaging modalities like echocardiography, such AI‑driven risk scores could become standard in cath‑lab workflows, guiding therapeutic intensity and follow‑up scheduling. Health systems may see reduced adverse events and lower costs through targeted interventions, while clinicians gain confidence from data‑backed prognostication. Ongoing multicenter trials will be crucial to confirm generalizability and to integrate these tools into existing clinical pathways.

AI-powered MRI evaluations predict STEMI outcomes better than existing risk scores

Dave Fornell · February 17, 2026

An advanced machine‑learning (ML) model that evaluates cardiac MRI results and clinical data predicts long‑term outcomes better than traditional risk‑stratification models in patients with ST‑segment elevation myocardial infarction (STEMI), according to a new study published in Radiology.[1] The analysis shows how artificial intelligence can improve individual patient assessments; this could become a key tool for improving care in the coming years.

“The major finding of our study was that the ML model demonstrated excellent predictive performance and strong discrimination for time to major adverse cardiovascular event (MACE) in the external test set,” wrote lead author Wei‑Hui Xie, MD, PhD, of the Department of Radiology at Renji Hospital in China, and colleagues.

Current risk‑stratification models do not incorporate a broad range of patient‑data parameters to predict MACE in STEMI patients. STEMI patients face a higher rate of cardiovascular death, recurrent heart attacks, unplanned coronary revascularization, stroke and rehospitalization for heart failure or arrhythmia. The study looked at combining more data and MRI imaging to create a personalized risk assessment for each patient and then compared the result to the popular Global Registry of Acute Coronary Events and the Thrombolysis in Myocardial Infarction scores.

Researchers initially included 67 variables to inform the AI model. The final model included established clinical predictors combined with features selected using recursive feature elimination. The study included 1,066 STEMI patients, with 682 in an ML training set and 384 in the final test set. The cohort was mostly men (904). During a median follow‑up of 40 months, 142 patients in the training set and 81 in the external test set experienced a MACE event.

The authors wrote that the ML model consistently demonstrated excellent discriminative ability and achieved the highest integrated area under the curve (AUC) across all major outcomes. The ML model showed the highest discrimination and lowest prediction error across all end points and consistently higher performance throughout the entire follow‑up period.

Although the ML algorithm demonstrated excellent predictive performance compared with existing clinical risk scores, the authors also noted limitations and the need to expand testing to new populations. This includes testing the model on cohorts outside of China and using more accessible imaging modalities, such as echocardiography.

---

Reference

1. Wei‑Hui Xie, Ruo‑Yang Shi, Jin‑Yi Xiang, Bing‑Hua Chen, et al. Machine Learning Using Clinical and Cardiac MRI Features to Predict Long‑term Outcomes in Acute STEMI. Radiology. Feb 17 2026; Volume 318, Number 2. DOI: https://doi.org/10.1148/radiol.251490