Based on Machine Learning, PRAISE Score Balances Risks in Post-ACS Care

The score is the first of its kind to be developed and validated in a cohort of only ACS patients.

Based on Machine Learning, PRAISE Score Balances Risks in Post-ACS Care

A new risk score based on machine learning shows promise in predicting death, MI, and bleeding following an ACS event, potentially allowing clinicians to better individualize care.

The score, called PRAISE, is not the first score proposed to aide in outcome prediction and dual antiplatelet therapy (DAPT) duration decisions for this cohort of patients; one specifically designed for use in patients at high bleeding risk was published just last week. But the machine learning component is novel.

“The study showed that a machine learning approach is feasible and effective and more accurate than currently existing predictive scores that are based upon conventional and classical statistical methods,” study co-author Ovidio De Filippo, MD (University of Turin, Italy), told TCTMD. “We believe that this is an important step toward personalizing and having a tailored approach for patients suffering from acute coronary syndrome in order to predict their risk of death in the year following the event, and also have a tailored approach to antithrombotic therapies.”

PRAISE’s current “most important application is the prognostication of events,” study co-author Guglielmo Gallone, MD (University of Turin), added. “We would like to use this score to predict, on one side, ischemic events and, on the other side, the bleeding [risk] of the patient. It is emerging that it's always more important to consider both parts in the same patient since the trade-off of a certain antiplatelet strategy in terms of aggression and potency is really dependent on both these components.”

Their findings were published in the January 16, 2021 issue of the Lancet.

Internal, External Validation

To train and validate their PRAISE model, De Filippo and Gallone along with lead author Fabrizio D’Ascenzo, MD (University of Turin), and colleagues used 1-year postdischarge all-cause death, MI, and major bleeding (BARC types 3 or 5) outcomes as well as 25 clinical features from 19,826 ACS patients in two registries. They also externally validated their model with 3,444 collective ACS patients from one randomized trial and three prospective registries.

Prediction was good for all three endpoints studied in both the internal and external validation cohorts.

One-Year Validation Outcomes

 

Area Under the Curve

95% CI

All-Cause Death

 

 

    Internal

0.82

0.78-0.85

    External

0.92

0.90-0.93

MI

 

 

    Internal

0.74

0.70-0.78

    External

0.81

0.76-0.85

Major Bleeding

 

 

    Internal

0.70

0.66-0.75

    External

0.86

0.82-0.89

 

Two-year outcome prediction was similar for each of the endpoints in the external validation population.

The PRAISE score also seemed to show equitable predictive value across low-, intermediate-, and high-risk patients for death, MI, and bleeding alike. “A gradual and progressive increase in absolute event rates was observed across risk classes for all the PRAISE scores,” the authors write, noting that intermediate- and high-risk individuals were more likely to report events than low-risk patients (P < 0.0001).

Bleeding tended to be most predominant in those at low risk of MI. Of the 2,327 patients deemed to be at high risk of major bleeding, 71.6% showed a low-to-moderate risk of MI. Similarly, among the 2,327 patients at high risk of MI, 71.6% showed a low-to-moderate risk of major bleeding.

“We believe that this paper lays the foundation for a clinical trial in which patients will be treated according to PRAISE-specified treatment strategies and randomized against clinical guideline-based attribution of dual antiplatelet therapies,” Gallone said.

De Filippo added: “This tool, [which] can be calculated in a very easy way, will be a useful tool for clinicians to [use] at follow-up for a full picture of the patient’s risk and to support clinical decision-making.”

Compared with other scores, De Filippo said PRAISE is the first to be able to “capture nonlinear relationships and variables.” Additionally, “our score is probably the first one that focuses on the patients just treated for acute coronary syndrome, whereas existing scores—DAPT, PRECISE-DAPT, and PARIS scores—were mostly derived from a selected cohort where patients who suffer from an acute coronary syndrome have a higher risk of both thrombotic and ischemic events that is higher in the first year after the first myocardial infarction and remains stable [and] higher thereafter. So selecting a cohort made just of patients who suffered from an acute coronary syndrome we believe is a strength.”

The study authors say they would like to see their model further externally validated in a randomized setting, perhaps according to high or low risk scores and differing DAPT durations.

‘Simple, Intuitive, and Easy to Implement’

In an accompanying editorial, Christian Templin, MD, and Davide Di Vece, MD (University Hospital Zurich, Switzerland), write: “Although new clinical scores are being continuously developed, few have proven their validity and effectiveness. . . . Compared with previous scoring approaches, the PRAISE score is based on one of the largest study populations and shows superior performance when externally validated. The variables included in the score are easily accessible from patient discharge information. Therefore, the score is simple, intuitive, and easy to implement in everyday clinical practice, also thanks to the PRAISE score calculator available online.”

The editorialists also highlight the researchers’ use of only ACS patients, and not stable CAD patients, to train and validate the PRAISE model. “[This] should be considered as an additional strength,” they write. “Although the ability of machine learning-based models to overcome limitations of traditional regression-based risk prediction systems remains a subject of debate, D’Ascenzo and colleagues’ study does not allow conclusions to be drawn in this regard, since a comparison with conventional statistical models is lacking.”

The PRAISE model represents a "considerable step forwards in enhancing the risk stratification of patients with ACS,” Templin and Di Vece conclude. Caveats, however, include potential confounding due to the observational nature of the derivation cohorts and the fact that the score was mainly validated in Italian patients. Additionally, since event prediction rests on clinical features, such as LVEF and age, that are linked to both ischemia and bleeding events, “the model does not always lead to therapeutic decisions that ultimately improve prognosis,” they say.

Disclosures
  • D’Ascenzo, De Filippo, Gallone, Templin, and Di Vece report no relevant conflicts of interest.

Comments