Artificial intelligence to improve ischemia prediction in rubidium positron emission tomography

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background Patients are referred to coronary artery disease (CAD) testing based on their pre-test probability (PTP). The available, easy-to-use risk prediction tools use three variables only and their diagnostic accuracy is limited. Artificial intelligence-based tools incorporate multiple variables and could improve PTP assessment due to the resulting combinatorial power. The factors in the AI model are not seen as independent individual values, but are recognized as patterns derived by a combinatorial analysis of the individual metabolic profile of each patient. Purpose We aimed to compare the diagnostic accuracy of a novel, artificial intelligence-based tool with commonly used PTP tools to predict ischemia. Methods Consecutive patients (n = 2417) referred for Rubidium-Position Emission Tomography (PET) were evaluated. PTP was calculated using the ESC 2013/2019 and ACC 2012/2021 guidelines, and a memetic pattern-based algorithm (MPA) model incorporating symptoms, vitals, ECG and laboratory findings. Five risk categories (very low to very high risk) were defined (<5%, 5–15%, 15–50%, 50–85%, >85%). Ischemia was defined as summed difference score (SDS) ≥2 on PET. Receiver operator characteristics were calculated and compared using the DeLong method. Results Known CAD, ischemia and scar (detected on PET) were present in 46.3%, 37.1% and 24.6%, respectively. The MPA model was most accurate to assess PTP of ischemia (AUC: 0.758, ESC 2013: 0.661, ESC 2019: 0.673, ACC 2012: 0.585, ACC 2021: 0.667, p<0.0001 each) as depicted in figure 1. The MPA’s sensitivity and negative predictive value to rule-out ischemia were 99.1% and 96.4%, respectively. The model allocated patients more evenly across risk groups, reduced the proportion of patients in the 15–85% range by 29% (ACC 2012) to 51% (ESC 2013), and was the only risk tool to correctly estimate ischemia prevalence in the very low risk group (table 1). Conclusion The MPA model significantly improved PTP assessment of ischemia. The MPA model has originally been derived on the endpoint coronary stenosis >50% and validated in three clinical studies (1–3). This study now confirms that the MPA model applied to the endpoint ischemia outperforms the current risk assessment tools and is most accurate to assess PTP. It enables clinicians to safely exclude ischemia based on readily available variables without advanced testing. It could be used to improve risk stratification of patients and to significantly reduce unnecessary non-invasive functional tests.