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

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Pulmonary hypertension (PH) is a serious condition often diagnosed late due to limited access to expert centers.
• Artificial intelligence (AI) offers potential for automated PH diagnosis through clinical decision support tools.
• Analysis of right heart catheterization (RHC) data is crucial for AI-based PH diagnosis but has not been previously reported.

Purpose of the Study:

• To develop and validate an automated method for PH diagnosis using AI and RHC data.
• To assess the feasibility of using large language models (LLMs) and machine learning (ML) for RHC data extraction and interpretation.
• To address the challenge of underdiagnosis of PH in a large patient population.

Main Methods:

• Retrospective cohort analysis of RHC studies from a statewide clinical network (UMMS) between January 1, 2016, and December 31, 2024.
• Development of a large language model (LLM)-driven Pattern Repository (LDPR) with three agents for unstructured RHC data extraction, validated by PH experts.
• Utilized machine learning (ML) to create predictive formulas for mean pulmonary artery pressure (mPAP) from systolic (sPAP) and diastolic (dPAP) pressures to handle missing data.

Main Results:

• The study included 11,029 patients and 17,292 RHC reports; LLM extraction precision for mPAP, sPAP, and dPAP was over 99.4%.
• A novel ML-derived linear equation (mPAP=1.51+0.43*sPAP+0.45*dPAP) achieved an R² of 0.94, outperforming existing methods.
• Applying the ML formula to 507 patients with missing mPAP identified 382 (75.3%) with PH, reclassifying them from no diagnosis.

Conclusions:

• A combined LLM-ML approach effectively automates PH diagnosis using RHC data in a large, diverse patient cohort.
• This method presents an efficient and scalable solution to combat PH underdiagnosis.
• Demonstrates the potential of generative AI in creating clinically actionable tools for cardiovascular disease phenotyping and diagnosis.