Resting Clinical and Summarized PET Variables for Prescreening Reduced Coronary Flow Reserve: A Leakage-Controlled Single-Center Feasibility Study

<title>Abstract</title> <p>Background Coronary flow reserve (CFR) derived from stress PET provides clinically important physiologic information, but a practical question remains whether variables available at rest can support low-CFR prescreening in selected clinical settings. This question should be distinguished from using resting information as a substitute for standard stress-based physiologic assessment. Methods We performed a single-center retrospective feasibility study of 354 patients who underwent cardiac PET with quantifiable CFR. Candidate predictors were restricted to resting clinical variables, resting hemodynamics, and summarized gated/rest PET parameters. All stress variables, myocardial blood flow measurements, and CFR-derived variables were excluded before modeling to prevent information leakage. We evaluated continuous CFR prediction and binary classification for reduced CFR using repeated internal cross-validation. The primary prescreening endpoint was CFR &lt; 2.0, and CFR &lt; 2.5 was assessed as a secondary endpoint. To prioritize a stricter estimate of transportability, we additionally performed internal temporal validation using a fixed 2021-01-01 split, model-block comparison, and threshold-stability analyses based on bootstrap confidence intervals and repeated nested resampling. Results In internal temporal validation, which was treated as the primary analysis, discrimination for the primary endpoint was weak: test area under the receiver operating characteristic curve (AUC) was 0.515 for clinical variables alone, 0.589 after adding resting hemodynamics, and 0.599 after adding summarized imaging variables. For CFR &lt; 2.5, the corresponding temporal test AUCs were 0.519, 0.608, and 0.657, indicating only limited discrimination. For CFR &lt; 2.0, the apparent training-derived threshold targeting sensitivity ≥ 0.85 did not transport well, with temporal test sensitivity 0.200 and a bootstrap 95% interval of 0.000-0.500. For CFR &lt; 2.5, temporal test sensitivity and specificity were 0.581 and 0.588, respectively. Repeated nested resampling showed broad threshold distributions and marked sensitivity-specificity trade-offs for both endpoints, indicating that uncertainty could be characterized more clearly but not converted into a stable screening cut-off. No stable fixed threshold was identified in the current dataset. Internal repeated cross-validation showed weak continuous CFR prediction (best out-of-fold performance: R = 0.275, MAE = 0.640, RMSE = 0.908) and only modest discrimination overall. Conclusions Resting clinical and summarized imaging variables did not provide performance consistent with a substitute for measured CFR or standard stress PET assessment. They may provide limited value for low-CFR prescreening or risk stratification, but the signal was weak for CFR &lt; 2.0 and only modest for CFR &lt; 2.5. No stable fixed threshold was identified in the current dataset. The main contribution of this study is therefore not a deployable rest-only model, but a leakage-controlled benchmark that clarifies the clinical boundary between rest-only prescreening and physiologic measurement.</p>