Combination AI-Machine Learning to Diagnose Pulmonary Hypertension: A Real-World Evidence Cohort Study

BACKGROUND Pulmonary hypertension (PH) is a highly morbid disease, but underdiagnosis is common outside of expert referral centers. Consequentially, there may be opportunities to automate PH diagnosis using artificial intelligence (AI) clinical decision support tools. Analysis of patient-level right heart catheterization (RHC) data is required to optimize AI-based PH diagnosis but has not been reported previously.

We performed a retrospective cohort analysis of all RHC studies (January 1, 2016 to December 31, 2024) performed at the University of Maryland Medical System (UMMS), which is a Maryland statewide clinical network of 12 hospitals serving >2 million patients. We developed an automated large language model (LLM)-driven Pattern Repository (LDPR) method, featuring three task-specific LLM agents for extracting unstructured RHC data, which was manually cross-validated independently by two PH experts. To address data missingness, we used machine-learning to develop formulae to calculate mean pulmonary artery pressure (mPAP) from systolic (sPAP) and diastolic (dPAP) PAP, using an 80/20 train-test split.

The study cohort included N=11,029 unique patients and 17,292 RHC reports (age 66±13.5 years; 43% female; 65% White, 30% Black or African American; mPAP, 28±11mmHg; 26% congestive heart failure). The precision for accurate mPAP, sPAP, and dPAP extraction by the LLM was 99.6%, 99.4%, and 99.4%, respectively, with a detection failure of 0.4%. A missing mPAP was noted in N=548 cases and N=507 unique patients (3.2% and 4.6%, respectively). When applying ML to the dataset, the simple, linear equation: mPAP=1.51+0.43*sPAP+0.45*dPAP returned the highest R2 of 0.94 and lowest mean square error of 8.3 mmHg, which outperformed linear equations used currently (all p20mmHg, and therefore reclassifying patients from no diagnosis to a diagnosis of PH.

In this retrospective cohort analysis, combination LLM-ML-based extraction and interpretation of RHC was used to automate PH diagnosis in a large and heterogenous patient population. This approach is an efficient and scalable solution to preventing under-diagnosis of PH and demonstrates the feasibility of generative AI for advancing clinically-actionable tools that can improve cardiovascular disease phenotyping and diagnosis in real-world settings. Competing Interest Statement S.M.S, M.E.M., S.C., G.R. and C.M.E. report no conflicts. K.Z. reports research grants from United Therapeutics and Cardiovascular Medical Research and Education Fund (CMREF) and advisory board fees from AstraZeneca. B.A.M. has received grants from Deerfield Company; has a patent PCT/US2019/ 059890 pending to None, a patent #9,605,047 issued to None, a patent PCT/US2020/066886 pending to None, and a patent BWH 2023/152/29618/0438P02 pending to None; has received 5R01HL139613/03 grant, National Institutes of Health grant R01HL163960, National Institutes of Health grant R01HL153502, and National Institutes of Health grant R01HL155096/01. Funding Statement The University of Maryland-Institute for Health Computing is supported by funding from Montgomery County, Maryland and The University of Maryland Strategic Partnership: MPowering the State, a formal collaboration between the University of Maryland, College Park, the University of Maryland, Baltimore, and the University of Maryland Medical System. Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The study was approved by the University of Maryland Institutional Review Board and requirement for informed consent was waived. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes Footnotes Funding: The University of Maryland-Institute for Health Computing is supported by funding from Montgomery County, Maryland and The University of Maryland Strategic Partnership: MPowering the State, a formal collaboration between the University of Maryland, College Park, the University of Maryland, Baltimore, and the University of Maryland Medical System. Conflict of interest: S.M.S, M.E.M., S.C., G.R. and C.M.E. report no conflicts. K.Z. reports research grants from United Therapeutics and Cardiovascular Medical Research and Education Fund (CMREF) and advisory board fees from AstraZeneca. B.A.M. has received grants from Deerfield Company; has a patent PCT/US2019/ 059890 pending to None, a patent #9,605,047 issued to None, a patent PCT/US2020/066886 pending to None, and a patent BWH 2023–152-29618-0438P02 pending to None; has received 5R01HL139613-03 grant, National Institutes of Health grant R01HL163960, National Institutes of Health grant R01HL153502, and National Institutes of Health grant R01HL155096-01. Data Availability All data produced in the present study are available upon reasonable request to the authors and after ethical approval.