USE OF MACHINE-LEARNING MODELS TO IDENTIFY CLINICAL FEATURES IN PATIENTS WITH PULMONARY ARTERIAL HYPERTENSION ASSOCIATED WITH A FUTURE CLINICAL WORSENING EVENT

SESSION TITLE: Pulmonary Vascular Disease Posters 2 SESSION TYPE: Original Investigation Posters PRESENTED ON: 10/10/2023 12:00 pm - 12:45 pm PURPOSE: Patients with pulmonary arterial hypertension (PAH) may decompensate rapidly with clinical worsening leading to right-heart failure or death. Better risk prediction for clinical worsening may facilitate early intervention to delay PAH progression. We used machine-learning models to identify clinical features present in PAH patients before a clinical worsening event. METHODS: Data were from the Mayo Clinic’s electronic health records (EHR). Eligible patients had PAH confirmed by right heart catheterization (RHC; mean pulmonary arterial pressure ≥25 mmHg; conducted 2015–2019), with either PAH medication (including parenteral prostanoids) recorded after the diagnosis date or a pulmonary hypertension (PH) International Classification of Diseases (ICD; v.9/10) code recorded ≤6 months of RHC. Patients with PAH medication use before diagnosis and chronic thromboembolic PH were ineligible. Clinical worsening events (all-cause hospitalization, death, parenteral prostanoid initiation, lung or heart/lung transplant, or worsening functional class) occurring ≥1 month after PAH diagnosis date were captured. Time windows of interest were labeled “progression-free” (1-month period prior to ≥12 months with no clinical worsening event) or “progressing” (1-month period immediately prior to clinical worsening event). A progressing window could overlap the 1-month period following diagnosis but could not include the diagnosis date. Machine-learning models (LASSO regularized logistic regression and XGBoost) were trained to classify between progressing and progression-free windows. Clinical features indicative of PAH clinical worsening were defined as those having a positive (associated with progression), statistically significant coefficient by logistic regression in the best-performing model. RESULTS: The association between clinical worsening events and features recorded in EHR ≤1 month of diagnosis (baseline) or the 1-month window of interest (progressing or progression-free) were analyzed in 2442 PAH patients (1023 with an event). Of 99 input clinical features, those most strongly associated with a future clinical worsening event (logistic regression; ranked highest to lowest) were: ICD code for chronic respiratory disease, dyspnea symptoms, chest pain symptoms, primary care physician visit and cardiologist visit within the progressing window; male gender, connective tissue disease associated PAH, and high erythrocyte distribution width value at baseline. CONCLUSIONS: Machine learning models using EHR data identified features associated with a future clinical worsening event in PAH patients. CLINICAL IMPLICATIONS: We show the potential to use routine clinical features to better identify PAH patients at risk of a clinical worsening event. Improved monitoring of patients at risk of PAH clinical worsening could aid early clinical intervention to optimize patient outcomes. DISCLOSURES: Employee relationship with nference Please note: Nov 2019 - present Added 03/30/2023 by Katherine Carlson, source=Web Response, value=Salary Employee relationship with nference, Inc. Please note: Dec. 2019-present Added 03/30/2023 by Corinne Carpenter, source=Web Response, value=Salary Employee relationship with nference Please note: june22-march23 Added 03/31/2023 by Deeksha Doddahonnaiah, source=Web Response, value=Consulting fee Consultant relationship with Janssen Please note: 2020-2022 Added 12/03/2022 by Hilary Dubrock, value=Consulting fee Grant/researchsupport relationship with Bayer Please note: 2021-2023 Added 03/13/2023 by Hilary Dubrock, source=Web Response, value=Grant/Research Employee relationship with Janssen Scientific Affairs, LLC Please note: 12/13/2021-curr Added 03/24/2023 by Hayley Germack, source=Web Response, value=Salary Stock Holder relationship with Johnson and Johnson Please note: 12/13/2021-curr Added 03/24/2023 by Hayley Germack, source=Web Response, value=Stock Holder Employee relationship with nference Please note: May 2019 - Present Added 03/31/2023 by Eli Silvert, source=Web Response, value=Salary No relevant relationships by Xiaoqin Tang Employee relationship with Johnson and Johnson (Janssen Pharmaceuticals) Please note: 2 years Added 03/31/2023 by Tobore Tobore, source=Web Response, value=Salary Employee relationship with nference, Inc. Please note: 2018-Present Added 03/30/2023 by Tyler Wagner, source=Web Response, value=Salary Employee relationship with Anumana, Inc. Please note: 2021-Present Added 03/30/2023 by Tyler Wagner, source=Web Response, value=Stock