Artificial intelligence in pulmonary hypertension: a systematic review

Pulmonary hypertension (PH) is characterized by elevated pulmonary pressures and right ventricular strain. Pulmonary arterial hypertension (PAH), a subtype, has a poor prognosis, especially when diagnosis is delayed. Artificial intelligence (AI) methods, including machine learning (ML) and deep learning (DL), offer potential for non-invasive prediction and risk stratification.

This systematic review assesses ML and DL applications for non-invasive diagnosis, classification, and prognostication in PH and PAH, with emphasis on methodological quality and clinical applicability.

A PRISMA-guided search identified studies using ML or DL on non-invasive clinical, imaging, or biomarker data, including omics and laboratory parameters. Study characteristics and heterogeneity were synthesized using the SWiM framework. Risk of bias was assessed using PROBAST+AI across participant selection, predictors, outcomes, and analysis.

Fifty-three studies were included. Most used clinical, echocardiographic, imaging, or molecular data. AUC values ranged from 0.71 to 1.00. DL approaches, especially convolutional neural networks, were increasingly applied but seldom externally validated. Nine studies were multicenter, four prospective, one combined retrospective and prospective cohorts, none were randomized controlled trials. The rest were retrospective single-center studies. In 15 studies, right heart catheterization was either not performed or not clearly reported. SWiM analysis showed substantial heterogeneity in study design and outcome definitions. According to PROBAST +AI, 44 studies (83%) had low risk of bias, though applicability concerns were common.

ML and DL models show promise for PH and PAH diagnosis and prognosis, but limitations in subclass differentiation, methodological transparency, and validation must be addressed in future research.