The power of data-driven ASSISTance in personalized testing for coronary artery disease

Introduction

Nearly 200 million people globally suffer from coronary artery disease (CAD),¹ half of whom initially present with chest pain.^1,2 Current guidelines emphasize the use of non-invasive investigation (Class 1 recommendation, Level of Evidence B) for suspected CAD.^3,4 Consequently, the use of functional stress tests and anatomical imaging through coronary computing tomography angiography (CTCA) for patients at low to intermediate risk of CAD has increased.^3,5 It is now recognized that, as well as providing diagnostic information, these imaging modalities can improve outcomes by enabling the implementation of preventative therapies.⁶ Whether to decide between a functional and anatomical test is a complex process incorporating multiple variables, and prone to the biases of an individual clinician. Such biases may lead to decision-making that leads to poorer outcomes for patients or disparities between patient groups.⁷ Data-driven machine learning may be able to identify where testing strategies differ from the optimal approach, thus reducing disparities, and achieving improvement in long-term cardiovascular outcomes.

Discussion

The linked study published by Oikonomou and colleagues in this issue of the journal explores cardiovascular outcomes of stable chest pain patients when comparing testing strategies employed in real-world care with the strategies recommended by the Application of Anatomical vs. Stress teSting decision Support Tool (ASSIST).⁸ The authors previously developed ASSIST to predict personalized benefit of anatomical vs. functional testing through phenomapping and an extreme gradient boosting algorithm,⁹ using participant level data from the PROMISE (Prospective Multicentre Imaging Study for Evaluation of Chest Pain) trial, with external validation in the SCOT-HEART (Scottish Computed Tomography of the Heart) trial.^6,10 In trial populations, the testing strategy recommended by ASSIST was associated with improved outcomes, but this study seeks to address whether this is robust in real-world practice, where decisions are not made as a consequence of random allocation.

From 134 216 individuals in Yale health system [anatomical n = 4030 (3%), functional n = 130 196 (97%)] and 3901 individuals in UK Biobank [anatomical n = 581 (14.9%), functional n = 3320 (85.1%)], 11 391 (8.5%) and 484 (12.3%) experienced the primary composite outcome all-cause death and acute myocardial infarction over a median follow-up of 4.9 years and 5.4 years, respectively. ASSIST would have projected better outcomes from anatomical testing for 18% of individuals in the Yale health system and 21.2% of the UK Biobank cohort, thus in both cases in excess of real-world practice, with female sex and a history of diabetes significantly associated with a lower likelihood of undergoing ASSIST-recommended testing strategy. Patients who underwent testing concordant with ASSIST recommendation were less likely to experience the primary outcome [Yale health system: adjusted HR 0.81 (95% CI 0.77–0.85); UK Biobank: 0.74 (95% CI 0.60–0.90)] after propensity score adjustment including risk factors. But why should a difference in testing strategy result in better outcomes? In a mechanistic sub-study, the authors leverage PROMISE data to demonstrate that ASSIST-concordant testing with CTCA was associated with a greater frequency of identifying high risk anatomic disease, thus potentially enabling better medical prevention strategies.

There are limitations to the study. This real-world analysis was heavily skewed towards functional testing. That the dataset is skewed to such a great extent implies that factors beyond those that can be accounted for likely influenced clinical decision-making. The use of all-cause mortality in the primary endpoint (a decision based on the lack of granularity in the Yale health system data) means that some of the benefit observed from an ASSIST-guided strategy may not be mechanistically linked to testing strategy. Furthermore, the exact role of ASSIST in the clinical pathway is still to be determined. ASSIST allows the projection of a new participant to a simulated PROMISE trial setting to predict the strategy that would work best in that setting, and its role could range from quantifying human-led diversion from an algorithmic approach to test selection to informing clinician decision-making between the different testing strategies (Figure 1).

Figure 1.

Factors influencing selection of anatomical vs functional testing in chest pain evaluation.

With multiple challenges and disparities present in investigation of suspected CAD,¹¹ the use of data-driven support tools has the potential to improve the fairness of care and patient outcomes. The authors should be congratulated for using real-world data to identify disparities in testing strategies that may disadvantage certain individuals, and thus the value of incorporating (objective) data-driven methods to inform decision-making. These findings add weight to the notion that we now need a prospective evaluation of a data-driven-informed selection of testing strategy in patients presenting with stable chest pain.