AI in Pulmonary Hypertension Demonstrates Progress—with Caveats

Pulmonary hypertension (PH) is difficult to diagnose early. Symptoms such as exertional dyspnea and fatigue are common across a range of cardiopulmonary conditions, and delays of several years between symptom onset and diagnosis are common. Against that backdrop, artificial intelligence (AI) has generated growing interest as a tool that could help identify patients earlier, refine risk assessment, and support clinical decision-making. A new systematic review published in the European Journal of Medical Research takes stock of where the field stands today.

The review evaluated 53 studies published between 2016 and 2025 that applied machine learning (ML) or deep learning (DL) techniques to PH and pulmonary arterial hypertension (PAH). Rather than focusing on hemodynamic data from right heart catheterization, the studies largely used noninvasive tests, including electrocardiograms (ECGs), echocardiography, chest radiographs, computed tomography (CT), cardiac magnetic resonance imaging (MRI), and biomarkers.

Detection Leads the Way

Most of the published studies focused on diagnosis, and many reported strong discrimination. Across all studies, area under the curve (AUC) ranged from 0.71 to 1.00, depending on the dataset and clinical objective.

Diagnostic performance varied substantially across modalities and study designs. ECG-based algorithms consistently demonstrated strong discrimination, while multimodal and MRI-based prognostic models generally achieved the highest reported performance in smaller, more homogeneous cohorts. Chest radiography and CT-based approaches also showed promise for noninvasive screening, although accuracy varied across datasets and depended on the reference standard used.

Looking Beyond Diagnosis

A smaller group of studies explored prognostic applications. These models sought to predict outcomes such as mortality, disease progression, or survival using combinations of imaging findings, biomarkers, and clinical data.

Several approaches showed meaningful improvements over established risk tools. Cardiac MRI-based models improved mortality prediction in PAH, while ML models integrating biomarkers and clinical variables demonstrated strong performance for long-term survival estimation. These findings suggest AI may eventually support more individualized risk stratification, although most studies remain exploratory.

A Strong Signal, but Is the Field Ready?

One of the review's recurring concerns was the inconsistent characterization of pulmonary hypertension itself. Many studies did not clearly distinguish between PH subgroups using the current ESC/ERS classification, despite important differences in pathophysiology, treatment, and prognosis. As a result, models trained on heterogeneous populations may be less useful for guiding subtype-specific clinical decisions.

Methodological limitations extended beyond cohort definition. Nearly 91% of included studies were retrospective, and 83% were conducted at a single center. Fewer than 40% performed external validation, making it difficult to know whether reported performance would hold up in different patient populations or healthcare systems.

The quality of the diagnostic reference standard also varied. In 15 studies, right heart catheterization, the gold standard for PH diagnosis, was either not performed, not clearly reported, or replaced by echocardiography alone.

Another challenge is interpretability. Many DL models function as black boxes, producing predictions without clearly explaining which features drove the result. Clinicians may be reluctant to rely on algorithms that cannot be readily understood or interrogated, particularly in a disease as complex as PH.

The Road Ahead

AI clearly shows potential to improve noninvasive detection and risk assessment in PH. At the same time, most remain proof-of-concept tools rather than clinically validated solutions. Because PH subtypes differ substantially in underlying biology, treatment strategies, and outcomes, failure to distinguish them limits the clinical utility of many AI models. Broader external validation, standardized diagnostic definitions, prospective testing, and greater transparency will be necessary before AI can move from promising research to routine practice.

For now, the most realistic role for AI is as an adjunct to clinical expertise, helping clinicians recognize patterns that might otherwise be missed while established diagnostic pathways remain firmly in place.

Reference:

Kramer T, Kramer M, Hagist C, Spinler S. Artificial intelligence in pulmonary hypertension: a systematic review. Eur J Med Res. 2025;30(1):1225. doi:10.1186/s40001-025-03557-5