Embargo 13 November – 10:15 CET
Key takeaways
- Few tools exist to identify patients at high risk of death after treatment for a ST‑elevation myocardial infarction, a type of heart attack.
- In the PERISCOPE project, machine-learning models were developed that could accurately predict in-hospital mortality and six-month mortality across three diverse databases.
- Prospective studies are now warranted to test if the machine-learning models can identify high-risk patients in clinical practice and improve their outcomes.
Basel, Switzerland – 13 November 2026: Machine learning models were able to predict mortality in hospital and in the six months after a type of heart attack, according to results presented today at ESC Digital & AI Summit 2026 [1].
Every year, more than two million people worldwide develop ST‑elevation myocardial infarction (STEMI) [2], a type of heart attack caused by complete blockage of one or more coronary arteries. Percutaneous coronary intervention (PCI) is effective for many patients after STEMI, but for some, the risk of mortality remains high after treatment.
As described by presenter, Doctor Andrea Rafanelli from Amsterdam University Medical Centre (UMC), The Netherlands, there is an unmet need for a risk prediction tool that can accurately identify patients at high risk of mortality after PCI: “Existing guideline-recommended scores, such as GRACE 2.0, estimate in-hospital and six-month mortality at presentation, before PCI, rather than after the procedure. Post-PCI risk models, on the other hand, often rely on numerous variables or on variables that are not routinely collected, which limit their use in daily practice.”
The PERISCOPE (Patient Risk StratifiCation for Post‑STEMI Complications) project aimed to develop machine-learning models to predict mortality in hospital and during the first six months after PCI. “The main aim is to be able to accurately identify those patients who need additional treatment and closer monitoring after the intervention,” she said.
A set of logistic regression models was trained on the US PREMIER PINC AI Healthcare Database (n=38,489), independently tested on US TriNetX data (n=119,035) and externally validated using Amsterdam UMC electronic medical record data (n=6,776). Across these three databases, in-hospital mortality rates ranged from 6.6% to 9.9%, while six-month mortality rates ranged from 11.4% to 23.7%.
The researchers found that the PERISCOPE models outperformed GRACE 2.0 score for predicting in-hospital mortality and six-month mortality across the three databases. In each case, the lower bound 95% confidence interval of the area under the receiver operating characteristic curve observed with the PERISCOPE models exceeded the upper bound seen with the GRACE 2.0 score.
“The machine-learning models provided good discrimination of both in-hospital and six-month mortality,” highlighted Doctor Rafanelli. “One of the most encouraging aspects was that the models showed consistent performances across the three datasets, which were geographically and clinically diverse, suggesting good generalisability,” she noted. “The PERISCOPE models may offer an easy-to-implement system to help triage patients at high risk of future events but first, their impact on outcomes must be tested in more diverse populations and prospective studies. Future research should assess how well the models work when integrated into the care pathway and how to make them actionable. In other words, what we actually do with high-risk patients once we have identified them. That could mean keeping patients in hospital a little longer or seeing them more often after discharge. And ultimately, we need to know whether those changes improve outcomes. Finally, it would be worth expanding the scope beyond STEMI to other forms of acute coronary syndrome,” Doctor Rafanelli concluded.
ENDS