A24-10 A Machine Learning Framework for Early Recognition of Pulmonary Arterial Hypertension Using Electronic Health Record Data

Rationale Pulmonary arterial hypertension (PAH) is a rare and progressive disease, where the guideline-recommended diagnostic confirmation is via right heart catheterization (RHC). However, due to non-specific symptomology, patients often experience delays in diagnosis that potentially impact disease outcomes. We sought to develop a machine learning (ML) based approach for early and appropriate detection of PAH. Methods Three cohorts were derived from Mayo Clinic electronic health record (EHR): confirmed PAH (N = 1,169), No PH (N = 11,600) and Other PH (Group 2-5 PH, N = 817), based on 2022 ESC/ERS guidelines. The study index was defined as the earliest diagnosis dates for PH cohorts (PAH & Other PH) or the latest PH negative echocardiogram (ECHO) or RHC for the No-PH cohort. Patients included in the study had at least one symptom (dyspnea, fatigue, syncope, chest pain, or lower extremity swelling) within one year prior to the study index. A two-step sequential ML modelling strategy was applied: Step 1 differentiated PH (PAH + Other PH) from No PH and step 2 differentiated PAH from Other PH. Models were trained using a combination of structured data, unstructured clinical notes to detect symptoms using natural language processing and ECG waveforms captured one year prior to study index. Models were tested both with and without ECHO parameters. SHAP (SHapley Additive explanations) values were used to determine important features contributing to models (without ECG waveforms). Results Models including ECHO parameters performed better compared to those without (Figure 1). The top feature impacting model without ECHO parameters was dyspnea in the preceding three months followed by heart failure and systolic blood pressure for the PH vs no PH; and T wave peak in V1 and ANA level in PAH vs other PH cohort. The top feature impacting model with ECHO was tricuspid regurgitation velocity followed by heart failure and right atrial pressure for the PH vs no PH; and mean blood pressure and left ventricular diameter in the PAH vs other PH cohort. Figure 1: Model performance in the two-step sequential approach PR-AUC=Precision Recall-Area Under Curve Conclusion This study demonstrates that ML models can aid in the identification of PAH in symptomatic patient populations. While ECG can identify patients for further testing, ECHO can further increase the accuracy and the need to refer patients to RHC for confirming diagnosis of PAH.