Prevalence-dependent thresholding and the safety of ruling out left main disease

This commentary refers to ‘Standard clinical and stress ECG variables to exclude left main and left main-equivalent disease: the MASTER study’, by M. De Carlo et al., https://doi.org/10.1093/eurheartj/ehag464 and the discussion piece ‘Prevalence-dependent thresholding and the safety of ruling out left main disease’, by M. De Carlo et al., https://doi.org/10.1093/eurheartj/ehag860. De Carlo et al. report a clinical and exercise electrocardiographic model intended to defer coronary angiography or coronary computed tomography angiography in patients with suspected chronic coronary syndromes. At an assumed 5% prevalence of left main coronary artery disease (LMCAD) and a 1:100 misclassification cost ratio, the negative predictive value was 98.2%.1 Their Discussion further states that a lower LMCAD prevalence would increase both sensitivity and negative predictive value, making adoption more advantageous. That interpretation is not supported by the supplementary analyses: after prevalence-dependent threshold selection, the higher negative predictive value at 0.9% prevalence was accompanied by lower sensitivity. At 5% prevalence, optimism-adjusted sensitivity and negative predictive value were 85.3% and 98.2% at the 1:100 ratio and 74.4% and 97.7% at 1:50.1 At 0.9% prevalence, Supplementary Table 7 reported sensitivities of 64.5% and 63.0%, with a negative predictive value of 99.5% under both ratios.1 The Mayo Clinic validation showed the same pattern: sensitivity fell from 87.2% to 79.5% at 1:100 and from 84.6% to 69.2% at 1:50, despite higher negative predictive values.1 These estimates imply that the proportion of affected patients missed increased from 14.7% to 35.5% at 1:100 and from 25.6% to 37.0% at 1:50. A lower prevalence reduces the absolute number of false negatives per 1000 patients, but it does not reduce the probability of missing LMCAD among patients who have the disease.