The main findings of this study are as follows: (1) The MPA model provided more accurate prediction of ischemia than the recommended PTP models (ESC 2013, ESC 2019, ACC 2012, ACC 2021). (2) The MPA model was the only model which correctly identified patients with a very low likelihood of ischemia. (3) The MPA model improved stratification across the whole PTP spectrum and reduced the proportion of patients in the intermediate range of 15–85% PTP by 28.9% (ACC 2012)–50.6% (ESC 2019). (4) The MPA model worked in patients without and with prior CAD, although it performed better in patients without prior CAD. Therefore, it should probably be used predominantly in patient cohorts without prior CAD. Hence, the MPA model is a useful tool to improve individualised assessment of pre-test probability and preselect patients for advanced cardiac testing. Furthermore, it could prevent patients with low probability of ischemia from unnecessary downstream tests, radiation exposure and costs. Therefore, it is a clear advancement in the direction of PPPM.

Current PTP tools are insufficient for patient preselection

Despite their easy use, the traditional risk prediction tools have two significant limitations. First, they either classify a substantial number of patients to have PTP < 15% with an insufficient sensitivity (80.8–90.7%) only (hence significantly underestimate the true prevalence), or they have an excellent sensitivity, but allocate a small proportion of patients in this PTP category only. Second, they allocate the majority of patients in the 15–85% range in which non-invasive imaging is recommended. Consequently, they are not useful in reducing the number of unnecessary non-invasive testing.

Comparison to earlier studies with the MPA model

The MPA model may overcome these issues to a clinically relevant extent with a more even distribution across PTP categories while maintaining an excellent sensitivity, NPV, NLR and FPR.

The MPA model’s overall AUC of 0.758 was good [24] and it performed clearly better compared to all other scores, also in the subgroup analyses. Still, the overall AUC was lower than reported in the earlier studies (original validation cohort Basel MPA 0.824 [20], LURIC validation 0.87 [17], Eurlings 0.87 [16]). This is most likely because the algorithm was trained and validated in previous works to detect the anatomic presence of CAD documented by invasive coronary angiography but not ischemia. In the present study, detection of ischemia by PET was used. A coronary vessel with an anatomic stenosis of > 50% as defined in the previous studies [20] does not necessarily translate into ischemia. In a sub-study of the COURAGE trial, Shaw et al. showed that approximately 40% of patients with at least one ≥ 70% stenosis had no or minimal ischemia only [25]. In the FAME trial, coronary stenoses in the range of 50–70% and 71–90% were not functionally significant in 65% and 20%, respectively [26]. Hence, this fact may explain at least in part the lower discriminatory power in the current study using the endpoint of ischemia, if compared directly to the initial MPA studies. Similar findings apply for the ACC and ESC scores [27, 28].

Performance of MPA model in subgroups

Despite the model being developed and trained in a cohort of patients without prior CAD, the MPA algorithm also performed acceptable in the subgroups (e.g. prior CAD). The AUC of each subgroup was lower than the AUC of the overall model, except for female patients where it was even slightly higher. The fact that both groups (with/without prior CAD) had worse AUC than the overall population is most likely because factors attributing for “prior CAD” significantly contribute to the model to estimate prevalence of CAD. The better AUC in female patients is probably because female patients had a lower prevalence of prior CAD.

Overall, the AUC of the MPA model was higher than all the PTP scores in each subgroup, highlighting the better discriminatory power and consistency of the test. Additionally, a higher MPA model score correlated well with the prevalence of ischemia. This may confirm the validity of this model also on a pathophysiological basis.

Potential field of application

A big advantage of the MPA algorithm is its ability to discriminate patients better across the whole spectrum of PTP, especially in the low- and very-low-risk categories. It exceeded the other models to correctly identify patients who have a very low prevalence of ischemia. If a certain cut-off for post-test probability was clinically accepted to abstain from testing (e.g. 5% or 10%, as proposed by certain authors [13]), this algorithm could be used to omit non-invasive testing in a significant number of patients.

The test characteristics to allocate patients in the very high-risk category (> 85%) were not as good as on the other side of the spectrum. This was most likely due to over-estimation of actual prevalence of ischemia, which was also observed with the other scores [27, 28]. This is most likely because all of them were developed and calibrated in cohorts where coronary artery disease was defined by luminal stenosis from an anatomical test (invasive angiography or computed tomography coronary angiography (CTCA)). As described above, significant luminal narrowing does not necessarily translate into ischemia. The clinical relevance of this slight overestimation of the prevalence of CAD appears insignificant since all of these high-risk patients need an advanced testing strategy anyway, be it non-invasive functional testing or an invasive angiogram. Hence, the MPA model is a better “rule-out” than “rule-in” test. Still, with the MPA’s false positive rate of 23.6%, this proportion is clearly below the prevalence of non-obstructed coronary arteries on routine angiograms as reported in certain cohorts (62.4%)29.

Comparison of study findings to published works

Miller et al. described a similar approach in a large multi-centre, international registry with > 20,000 patients [19]. They used patient specific data available prior to the scan and a machine learning–based algorithm to predict an abnormal myocardial perfusion [19]. The AUC to predict an abnormal scan was 0.762 (95% CI 0.750–0.774), which was similar to our MPA algorithm (0.758, 95% CI 0.739–0.777). Using their ultra-high sensitive threshold (which is approximately equivalent to our low PTP threshold (PTP < 15%)), test characteristics were comparable (sensitivity: 96% vs. 97%; NPV: 95% vs. 95%; 15.5% vs. 17.9% of patients below threshold). But, our very low PTP threshold exceeded the described ultra-sensitive threshold with a sensitivity of 99% and NPV of 96%. However, comparability is limited because Miller et al. included two variables in their model (prior CAD and past myocardial infarction) which account for the major part of the model. Even without including these two important factors, our model outperformed the described model if the < 5% cut-off is used. Furthermore, they did not include biomarkers and the endpoints differed significantly (SDS ≥ 2 on PET (this publication) vs. SSS ≥ 3 on SPECT (Miller)).

In another study, Ismaeel and colleagues compared an artificial neural network (ANN) with two older PTP tools (Diamond Forrester, Morise) to predict ischemia [18]. Similar to our study, the AI model outperformed the PTP tools and had a better discriminator power and good test characteristics to rule out ischemia (sensitivity 91%, negative predictive value 98%). The AUC of the ANN model was slightly lower than our MPA model (0.7 vs. 0.76). Comparability with this small study (n = 486) is difficult, because they used PTP tools which are not recommended anymore in the guidelines, the endpoint ischemia was not well defined and two different functional tests (SPECT, stress echocardiography) were used, which are less sensitive and less specific than PET.

Using the MPA model

The test characteristics of a given model strongly depend on the prevalence of the disease. Hence, cut-off points need to be adjusted depending on the cohort being tested. Therefore, two different calibrations of the MPA model are available [16, 17, 20] (MPA model and MPA model low risk; Table S5). Since these were calibrated in different cohorts with different prevalence of CAD, they use different cut-off points and must not be swapped interchangeably. As shown in Table S6, the MPA low-risk model [16] stratifies better in the high and very high-risk categories, but significantly underestimates ischemia prevalence in the low-risk category of the current study cohort.

Therefore, in order to ensure accurate risk stratification using the MPA model, it is important to select the appropriate cut-off points depending on the clinical setting of the patient (e.g. as used for risk stratification or screening in a general practitioner’s (GP) office vs. a diagnostic test in patients referred to a cardiologist’s office or hospital).

Despite better ischemia prediction in the very high PTP category (> 85%) compared to the other scores, the MPA model performed best to exclude ischemia. It could be used as a gatekeeper to reduce costs while maintaining its excellent test characteristics (cut-off > 15% PTP, sensitivity 97.3%, NPV 94.5%, NLR 0.099). Based on clinical information, biomarkers and ECG findings, it could be applied by primary care physicians to triage patients before they are referred for further downstream cardiac testing.

Limitations

Data from this project arise from a single centre. Images were analysed according to current guidelines by a small, steady and experienced team of Cardiologists and Nuclear Medicine Specialists reaching consensus. Hence, data interpretation was performed in a standardised and homogeneous way.

The four scores (ESC, ACC) were initially developed to assess pre-test probability of significant luminal stenosis in patients without prior CAD. We applied these scores in a mixed population with and without prior CAD which could limit the scores’ overall performance. However, we provide the subgroup analysis for both patient with and patients without prior CAD.