Abstract
BACKGROUND: Pulmonary hypertension (PH) is a pathophysiological problem that may involve several clinical symptoms and be linked to various respiratory and cardiovascular illnesses. Its diagnosis is made invasively by Right Cardiac Catheterization (RHC), which is difficult to perform routinely. Aim of the current study was to develop a Machine Learning (ML) algorithm based on the analysis of anamnestic data to predict the presence of an invasively measured PH. METHODS: 226 patients with clinical indication of RHC for suspected PH were enrolled between October 2017 and October 2020. All patients underwent a protocol of diagnostic techniques for PH according to the recommended guidelines. Machine learning (ML) approaches were considered to develop classifiers aiming to automatically detect patients affected by PH, based on the patient's characteristics, anamnestic data, and non-invasive parameters, transthoracic echocardiography (TTE) results and spirometry outcomes. RESULTS: Out of 51 variables of patients undergoing RHC collected, 12 resulted significantly different between patients who resulted positive and those who resulted negative at RHC. Among them 8 were selected and utilized to both train and validate an Elastic-Net Regularized Generalized Linear Model, from which a risk score was developed. The AUC of the identification model is of 83 % with an overall accuracy of 74 % [95 % CI (61 %, 84 %)], indicating very good discrimination between patients with and without the pathology. CONCLUSIONS: The PH-targeted ML models could streamline routine screening for PH, facilitating earlier identification and better RHC referrals.