Beyond detection: developing probabilistic hemodynamic phenotyping after a positive point-of-care pulmonary hypertension test

Background Pulmonary hypertension (PH) is a heterogeneous condition in which hemodynamic subtype directs management and specialist referral. PH can be identified non-invasively at the point of care by a previously developed, regulatory-cleared machine-learned diagnostic (area under the receiver operating characteristic curve 0.95, sensitivity 82%, specificity 92%), but a positive test result raises a second clinical question: which patients have pre-capillary disease and therefore warrant referral to a specialist PH center. A probabilistic post-test phenotyping model was developed to estimate hemodynamic subtype probabilities and support rule-out of pre-capillary PH among test-positive patients. Methods Subjects with right heart catheterization (RHC)-confirmed PH [mean pulmonary artery pressure (mPAP) ≥25 mmHg] enrolled in a prospective multisite trial (NCT04031989) who tested positive on the previously validated point-of-care PH diagnostic formed the analysis cohort ( n = 139). Hemodynamic subtypes were assigned per 2015 European Society of Cardiology/European Respiratory Society (ESC/ERS) criteria. A one-vs.-one (OvO) Gaussian mixture model (GMM) framework, combining unsupervised clustering with post hoc supervised assignment of subtype probabilities was applied to features derived from non-invasive orthogonal voltage gradient and photoplethysmographic signals to generate subject-level subtype probability estimates. Blinded validation in a large dataset has not yet been performed; development results are reported, supported by a small blinded independent dataset. Results At an exemplar post-hoc rule-out threshold of 16% pre-capillary probability, the GMM OvO approach yielded sensitivity 86% (95% CI: 71%–95%), specificity 51% (95% CI: 41%–61%), negative predictive value 88% (95% CI: 73%–96%) at an assumed pre-capillary prevalence of 34%, and a negative likelihood ratio (LR−) of 0.27 (95% CI: 0.07–0.70) on the development cohort. Supportive out-of-sample discrimination was observed in the blinded independent cohort. Conclusions A GMM OvO probabilistic model applied to non-invasive point-of-care signals can generate PH subtype probability estimates and may support rule-out of pre-capillary disease among true-positive patients. The model reuses signals already acquired for PH detection and can therefore be implemented as a software addition to an existing point-of-care workflow. Independent blinded validation is required to comprehensively characterize model performance.