Machine learning to detect pulmonary hypertension associated with chronic obstructive pulmonary disease (PH-COPD)

<bold>Background:</bold> PH is a key prognostic indicator in patients with COPD but can be difficult to diagnose due to non-specific symptoms. Right heart catheterization (RHC) is the gold standard diagnostic test for PH; however, it is invasive and not always performed for COPD patients. This study trained machine learning models to identify patients with PH-COPD, using electrocardiogram (ECG) and electronic health record (EHR) data. <bold>Methods:</bold> Of 99,970 adults with COPD between 2015-2019 at Mayo Clinic, 3,118 PH-COPD patients were found with 1) mean pulmonary arterial pressure (mPAP) >20 mmHg on RHC or tricuspid regurgitation velocity (TRV) >3.4 m/s via echocardiogram, 2) COPD diagnosis before or ≤3 months after PH diagnosis, and 3) no Group 1, 2, or 4 PH. A COPD only cohort of 6,248 patients without PH was identified with mPAP ≤20 mmHg on all RHCs and no TRV >3.4 m/s, or if no RHC, all TRV ≤2.8 m/s. Self-supervised machine learning models were developed using 2.4M patients’ EHRs and 9M ECGs from Mayo Clinic, and then trained to detect PH-COPD utilizing only ECGs or both ECG and EHR data. <bold>Results:</bold> Models performed equally well (area under the curve, sensitivity/specificity): ECG only 0.87, 79.8%/78.7%, both ECG and EHR 0.88, 81.5%/80.9%. These findings suggest that models can be leveraged in a clinical system to improve detection of PH among COPD patients. This could facilitate PH-COPD patient referrals to PH specialty centers for individualized care in accordance with current guidelines, potentially improving clinical outcomes.