Artificial Intelligence in Pulmonary Hypertension: Current State and Future Prospects

Abstract

Pulmonary hypertension (PH) remains diagnostically challenging due to the associated non-specific symptomatology and frequent diagnostic delays, both of which contribute to increased morbidity and mortality. Meanwhile, right heart catheterization is the diagnostic gold standard; nonetheless, the invasive nature and limited accessibility of this technique limit its routine use, particularly in resource-constrained settings. This review evaluates computational approaches that may enhance PH diagnosis through advanced analysis of cardiovascular imaging data. We conducted a comprehensive literature review focusing on computer-assisted diagnostic methods in PH across multiple imaging modalities, including electrocardiogram, chest X-ray, echocardiography, cardiac magnetic resonance (CMR) imaging, and cardiac computed tomography (CCT). Eligible studies were analyzed for diagnostic performance, clinical applicability, and methodological rigor. Preliminary studies have demonstrated promising performance in detecting early or subclinical PH phenotypes across various imaging platforms. Advanced imaging modalities benefit from automated segmentation and quantitative analysis, and CMR- and CT-based approaches demonstrate high diagnostic accuracy. Current artificial intelligence (AI) models face significant challenges related to clinical interpretability, external validation across diverse populations, and seamless integration into existing diagnostic workflows. Most studies are based on single-center retrospective cohorts, underscoring the need for multicenter prospective validation. To address these challenges, future research must prioritize advancing methodological transparency through explainable AI (XAI) and ensuring data privacy via federated learning. Crucially, the next phase of innovation lies in synergistic multimodal fusion (MMF) architectures that synthesize heterogeneous data—ranging from imaging to hemodynamics—to enhance phenotypic precision. Furthermore, leveraging large language models (LLMs) for computational phenotyping from electronic health records offers a scalable solution for identifying undiagnosed patients. Finally, realizing clinical translation requires rigorous multicenter prospective validation and seamless integration into existing workflows to ensure these tools effectively support decision-making in real-world practice. While computational approaches in PH diagnostics show promise for improving early detection and diagnostic accuracy, significant challenges remain before widespread clinical adoption. Future development should prioritize multicenter validation, standardized frameworks integrating multi-parametric imaging with clinical biomarkers, and transparent methodologies to support clinical decision-making. The integration of these computational tools into conventional diagnostic pathways may enhance PH management through earlier detection and more precise risk stratification.