Precision medicine gets an ASSIST in diagnostic testing

This editorial refers to ‘A phenomapping-derived tool to personalize the selection of anatomical vs. functional testing in evaluating chest pain (ASSIST)’, by E.K. Oikonomou et al., doi:10.1093/eurheartj/ehab223. Potential future decision pathway incorporating the range of risk for symptomatic patients with suspected coronary artery disease in whom testing is being considered. At each step, applying dedicated risk classification tools to guide management decisions in proper sequence should improve patient care. Potential future decision pathway incorporating the range of risk for symptomatic patients with suspected coronary artery disease in whom testing is being considered. At each step, applying dedicated risk classification tools to guide management decisions in proper sequence should improve patient care. Over 40 years ago, George Diamond and James Forrester published their landmark application of Bayesian analysis for determining pre-test probability in patients suspected of having obstructive coronary artery disease (CAD).1 Since that time, the concept of assessing pre-test probability to inform testing decisions and estimate clinical outcomes has been a central tenet of clinical cardiology. While multiple models have optimized identification of patients most (or least) likely to have obstructive CAD and future events,2,3 contemporary diagnostic approaches continue to suffer from low diagnostic yield and excess cost. Further, no currently available models provide guidance on which test to choose. For most clinicians, an important branch point in the decision tree is deciding between so-called ‘functional’ and ‘anatomical’ testing. Large clinical trials of unselected patients consistently demonstrate that preventive treatment is improved by initial anatomical testing, resulting in either comparable outcomes at initial functional testing (PROMISE; Prospective Multicenter Imaging Study for Evaluation of Chest Pain) or reduced death/myocardial infarction (SCOT-HEART; Scottish Computed Tomography of the Heart),4,5 but this pathway has not been universally adopted. This clinical uncertainty is reflected in the most recent guideline recommendations,3 and raises the question of whether the testing strategy should differ by patient groups with suspected CAD. Hypothesis-generating subgroup analyses from the PROMISE trial suggest that certain patient phenotypes may benefit more from one testing strategy vs. another. For example, coronary computed tomography angiography (cCTA) appeared to have greater prognostic value in women,6 diabetic patients,7 and younger patients,8 while functional stress testing was more predictive of future clinical outcomes in older patients.8 However, even if these findings are confirmed prospectively, converting them into testing strategy recommendations is problematic in heterogeneous populations with multiple subgroup memberships, each with conflicting benefit. For example, what is the optimal diagnostic approach in a patient with competing characteristics (such as diabetes and old age)? Given the importance of selecting the best test for the individual patient and the clinical equipoise across testing types, Oikonomou and colleagues, in their study published in this issue of the European Heart Journal, set out to determine if extensive phenotyping and advanced analytics could predict individual outcomes depending on selected test type.9 Utilizing patient-level data from two large randomized controlled trials (PROMISE and SCOT-HEART), including all available pre-testing data (57 variables in PROMISE), testing performed, and clinical outcomes, the investigators computed a dissimilarity index that identified neighbourhoods including the 5% most phenotypically similar participants. The authors found that the adverse event risk varied between testing type (anatomical vs. functional) by phenotypic neighbourhood. From there, the authors utilized machine learning methods to derive ASSIST (Anatomical versus Stress teSting decIsion Support Tool), a parsimonious model trained on the highest-yield 12 commonly available clinical features from the original group of 57. What resulted is a novel contribution to the armamentarium of clinicians routinely caring for patients with suspected CAD. While the use of precision techniques (including risk scores) to guide management decisions is not new to cardiology, ASSIST brings additional rigour to guide diagnostic test choice by using a comprehensive, phenomapping-derived, decision support tool. The strengths of the study include the use of large randomized trial data with dedicated follow-up and adjudicated outcomes, use of sophisticated machine learning methods, appropriate verification of results with sensitivity analyses, and the creation of a straightforward tool suitable for clinical practice. Yet, before ASSIST can be implemented into widespread clinical use, there are some important considerations. The derivation and validation of ASSIST utilize available data from PROMISE and SCOT-HEART, which assessed functional testing vs. cCTA. However, both functional and anatomical testing methods have progressed significantly since the collection of those data. For example, the introduction of non-invasive computationally derived fractional flow reserve computed tomography (FFRCT) technology adds haemodynamic information to cCTA. Data from the PROMISE trial suggest that the addition of FFRCT to cCTA alone would reduce both overall catheterization rates (from 12.2% to 7.8%) and the rate of catheterization without obstructive CAD (27% to 15%).10 In the observational Prospective LongitudinAl Trial of FFRCT: Outcome and Resource Impacts (PLATFORM),11 the investigators prospectively assessed a cCTA ± FFRCT strategy and demonstrated improved diagnostic yield and reduced cost. Similar advances have been made in the use of positron emission tomography, cardiac magnetic resonance, and other forms of functional testing over the past decade. The impact of ongoing innovation across non-invasive imaging modalities is an important consideration when applying potentially outdated risk models in contemporary practice. Ultimately, prospective validation of any precision strategy is required to determine its true value. Thus, a randomized clinical trial demonstrating the feasibility of implementation, clinical utility, and costs of using the risk model is essential. One of the first cardiovascular disease trials assessing the value of precision testing is now ongoing. The Prospective Randomized Trial of the Optimal Evaluation of Cardiac Symptoms and Revascularization (PRECISE) is a randomized trial enrolling ∼2100 participants that will compare a precision evaluation strategy utilizing the PROMISE Minimal Risk Tool with a usual care strategy in patients presenting with stable chest pain and suspected CAD (ClinicalTrials.gov identifier NCT03702244). The precision strategy will include guideline-recommended medical management without planned testing in participants determined to be low risk for CAD or events, and cCTA with selective FFRCT in participants at higher risk. The usual care arm will include non-invasive functional testing or invasive angiography at the clinician’s discretion. PRECISE will therefore implement a validated risk model into a diagnostic testing approach and rigorously evaluate the impact of such an approach on clinical outcomes, quality of life, and cost. A similar prospective implementation trial using ASSIST will be critical to demonstrate the feasibility of integrating the model into clinical practice and the impact of ASSIST. While precision strategies may improve the current testing paradigm, testing decisions will never be formulaic. Any approach is limited by numerous factors that influence the final shared decision between patient and clinician. Local expertise and test availability, individual contraindications to exercise or pharmacological stress testing, concurrent indications for thoracic imaging, suspected structural heart abnormalities, and radiation sensitivity are just some of the multitude of factors influencing these decisions,12 As always, patient preference is the final arbiter of any diagnostic testing decision. Aside from prospective validation, what are the next steps to use ASSIST in making diagnostic testing decisions? The prospect of machines providing real-time decision support to clinicians without requiring additional activation energy is a worthy goal, but often feels just beyond our grasp. Seamless integration into ‘real-world’ clinical workflow would be a high priority. While the authors have created a platform allowing for manual data input to generate a clinical recommendation, integration into the electronic health record (EHR) might provide even greater value. This would involve automatically pulling data from the EHR and providing ASSIST results to the clinician at the point of care. Interestingly, while Oikonomou and colleagues narrowed down their variable list from 57 to 12 to maximize usability for clinicians,9 fully automating this process with the EHR could remove barriers to a truly tailored risk classification, more closely resembling the full map of unique phenotypic neighbourhoods. An ideal, integrated view of the future of testing decisions encompasses the entire spectrum of symptomatic patients with suspected disease requiring diagnostic testing (Graphical Abstract). At each stage, applying dedicated risk classification tools that can guide management decisions in proper sequence will improve patient care, starting with clinical risk algorithms (pooled cohort equations or equivalent) to assist in preventive treatment decisions,13 followed by the 2019 European Society of Cardiology pre-test probability for the CAD algorithm,3 the PROMISE Minimal Risk Tool to identify symptomatic patients who may not require immediate testing,2 and the ASSIST model to direct patients needing testing to anatomical vs. functional testing.9 The final goal is clear: to improve clinical outcomes for all patients across the risk spectrum via the sequential application of precision medicine.