AI decision support tool did not improve outcomes among patients with suspected MI
Click Here to Manage Email Alerts
Key takeaways:
- AI to support clinical decision-making did not significantly improve outcomes in patients with suspected MI.
- Among all-comers in the ED, the AI tool began to show signs of potential benefit in CV outcomes.
An AI decision support tool embedded in routine care did not meaningfully improve outcomes for patients presenting to the ED with myocardial injury, but showed signs of benefit in an all-comer cohort, a speaker reported.
The results of the RAPIDxAI trial were presented at the European Society of Cardiology Congress.
“The purpose of this study was to evaluate in clinical care whether or not the availability of AI-based decision support aligned with the Fourth Universal Definition of MI led to superior care,” Derek P. Chew, MBBS, MPH, PhD, clinical and interventional cardiologist, cardiovascular health systems researcher and director of Monash Heart/Victorian Heart Hospital in Melbourne, Australia, said during a press conference. “More importantly, if it had an impact on CV death, myocardial infarction and unplanned revascularization in patients who presented to emergency departments and had myocardial injury.”
Using an established clinical registry, Chew and colleagues trained AI to phenotype patients with myocardial injury based on the Fourth Universal Definition of MI. The AI decision support tool was then redesigned to be embedded in real-time clinical workflow.
To evaluate the impact of the decision support tool on clinical outcomes, 12 hospitals were randomly assigned to the AI tool implemented in the ED or usual care.
In the primary analysis of 3,029 patients who presented with suspected MI, risk for the composite endpoint of CV death, MI and unplanned CV-related readmission was not significantly different between patients managed with the AI tool or usual care at 6 months (HR = 0.99; 95% CI, 0.86-1.14; P cluster = .872; win ratio = 1.04; 95% CI, 0.9-1.19; P = .606).
A separate, all-comer analysis including 14,131 patients was also conducted.
Among 5,466 patients with myocardial injury, use of the AI did not significantly impact risk for the primary composite endpoint compared with usual care (HR = 0.92; 95% CI, 0.82-1.04; P cluster = .74; win ratio = 1.09; 95% CI, 0.97-1.21; P = .156).
In the all-comer cohort, although not statistically significant, Chew and colleagues reported a trend toward lower risk for CV death, MI and unplanned CV-related readmission among those managed with the AI tool compared with usual care (HR = 0.8; 95% CI, 0.62-1.03; P cluster = .084; win ratio = 1.15; 95% CI, 1.04-1.27; P = .007).
“AI algorithms can improve the diagnosis of myocardial infarction and myocardial injury, but presenting this to clinicians did not translate to improvement in clinical outcomes within a cluster randomized trial embedded in clinical practice,” Chew said during the press conference. “AI remains the disruptive technology of our era, and the potential diagnostic insights it provides to nonexpert clinicians remains promising, but translation to clinical outcomes will require careful integration into the health care system.
“The value of these diagnostic insights needs to be continuously prospectively validated in clinical trials,” he said.
Collapse
Chew D. Hot line 12. Presented at: European Society of Cardiology Congress; Aug. 30-Sept. 2, 2024; London (hybrid meeting).
Click Here to Manage Email Alerts