Unsupervised Machine Learning to Identify Unique Clusters of Systemic Sclerosis Pulmonary Hypertension Patients

ABSTRACT Unsupervised Machine Learning to Identify Novel Clusters of Systemic Sclerosis Pulmonary Hypertension Patients Pulmonary hypertension (PH) is a common devastating complication of systemic sclerosis (SSc), an autoimmune condition characterized by fibrosis and inflammation of the skin and internal organs, particularly the lungs and blood vessels. SSc patients can have PH related to numerous etiologies, most notably pulmonary fibrosis/interstitial lung disease (ILD, termed group 3 PH) and/or a primary pulmonary vasculopathy (pulmonary arterial hypertension [PAH], group 1 PH). Since SSc patients have varying degrees of ILD and patients with ILD-PH and PAH can have identical pressures on right heart catheterization, it is challenging to precisely classify individual patients using our existing scheme (i.e., group 1 vs group 3). This has important implications for PH treatment choices and therefore further study is needed to improve our classification system. The Pulmonary Vascular Disease Phenomics Program (PVDOMICS) is an NHLBI-funded project aiming to deeply phenotype PH patients. The current project will be a secondary analysis of the PVDOMICS dataset and the Johns Hopkins PH Registry to accomplish the following Specific Aims: Aim 1: Use unsupervised machine learning to identify distinct clusters of SSc pre-capillary PH patients, incorporating echocardiography, cardiac MRI, invasive cardiopulmonary exercise testing, clinical characteristics, and biomarkers. Aim 2: Evaluate transplant-free survival and hospitalization rates in the newly identified clusters. Principal component analysis and K-means clustering will be used to determine the novel clusters, independent of our current classification scheme. Cox regression and receiver operating characteristics curve analysis will be used to determine if the newly discovered clusters differ based on clinically relevant outcomes and have improved discrimination and calibration compared to the traditional classification system. Successful completion of these aims will establish novel methods for phenotyping SSc-PH patients that transcend our current paradigm of group 1 vs group 3 classification. The potential clinical relevance of the cluster-based classification method in patient-centered outcomes will be established. The long-term objective of this project is to generate data to plan prospective phenotyping studies testing personalized treatment strategies based on SSc-PH cluster assignment, which has the potential to change the paradigm of how we classify and treat this complex subset of pulmonary hypertension patients.