Non-invasive diagnosis of post-capillary pulmonary hypertension: validation of a machine learning diagnostic tool in the PVDOMICS cohort

Advanced age and increasing comorbidities complicate the differentiation between group 1 and 2 pulmonary hypertension (PH) and increase the number of unnecessary right heart catheterizations. Therefore, a machine learning (ML) predictive model was developed to diagnose group 2 PH using 20 non-invasive parameters. We aimed to validate its performance in the PVDOMICS cohort. We extracted the required parameters from the eligible group 1 and 2 PH patients to perform the ML prediction. The results were compared to the registered diagnosis in the database, as well as the haemodynamic diagnosis based on the 2022 ESC/ERS guidelines for PH. The PVDOMICS cohort included 174 group 1 and 216 group 2 PH patients. According to the registered diagnosis, the diagnostic model identified group 2 PH with a specificity of 93%, positive predictive value (PPV) of 92%, and sensitivity of 67%. Accuracy was higher in patients with mPAP ≥25 mmHg, with a PPV of 93% for a sensitivity of 71%. According to the haemodynamic diagnosis, the model had a specificity of 95% and PPV of 97% for a sensitivity of 54%. The few group 1 patients inaccurately diagnosed as group 2 had remarkably many characteristics consistent with group 2 PH. Applied to this US cohort, the ML model had a slightly lower accuracy compared to the Belgian development and validation cohorts, which could be related to a higher prevalence of metabolic syndrome in the US cohort. The specificity and PPV of the model encourage further prospective validation.