Abstract 4365733: Validation of a Non-invasive Machine Learning Algorithm to Assess Elevated Pulmonary Capillary Wedge Pressure at Point-of-Care

Abstract

Introduction: Pulmonary capillary wedge pressure (PCWP), which reflects left ventricular filling pressure (LVFP), can only be reliably obtained using invasive right heart catheterization (RHC). Accurate assessment of PCWP is central to heart failure diagnosis and management, particularly with preserved ejection fraction (HFpEF), where LVFP is elevated despite normal EF. Although transthoracic echocardiography (TTE) is commonly used in clinical practice, it has limited sensitivity for detecting elevated LVFP. PCWP is also required to differentiate pulmonary hypertension (PH) subtypes, which have distinct treatment strategies. Clinicians are currently lacking an effective, non-invasive, point-of-care test for estimating PCWP, which could facilitate earlier diagnosis and guide appropriate treatment for patients with HFpEF, PH, and related conditions. Hypothesis: We previously developed an algorithm to estimate PCWP from non-invasive signals. We hypothesized that it would meet the co-primary endpoints of sensitivity >70% and specificity >60% to identify elevated PCWP (>18mmHg) in a blinded validation cohort. Methods: The test consists of a signal acquisition device that collects orthogonal voltage gradients using seven thoracic sensors, and a photoplethysmographic signal from a fingerclip. A total of 243 signal features are extracted and input into a static CatBoost model, producing a continuous score. The score is classified as test-positive or test-negative using a predefined and locked cutpoint. The validation cohort included subjects with elevated PCWP by RHC, as well as symptomatic subjects with TTEs showing normal diastolic function and low probability of PH based on published guidelines. Results: The validation cohort contained 108 PCWP-positive and 147 PCWP-negative subjects. The algorithm achieved a sensitivity of 82.4% (95% CI: 75.2–89.6%) and specificity of 83.0% (95% CI: 76.9–89.1%), with an AUC of 0.91 (95% CI: 0.89–0.93). These results exceeded the predefined performance goals (sensitivity: p=0.0005; specificity: p<0.0001). The positive and negative likelihood ratios were 4.85 and 0.21. Critically, no significant differences were observed across sex, age, and racial subgroups. Conclusions: The algorithm met the pre-defined endpoints in the independent blinded validation cohort, with robust ability to increase or decrease the odds of disease by approximately five-fold, supporting the clinical utility of the test in the assessment of PCWP elevation.