JMIR Publications

ABSTRACT

Background:

Pulmonary hypertension (PH) requires invasive hemodynamic confirmation, but right heart catheterization (RHC) is not always accessible. Noninvasive imaging is central to triage, yet the comparative diagnostic performance of conventional and artificial intelligence (AI)-enhanced modalities remains uncertain.

Objective:

To compare the diagnostic accuracy of conventional and AI-enhanced noninvasive imaging modalities for PH using RHC as the reference standard and to determine whether AI-enhanced approaches provide complementary value for clinical triage.

Methods:

In this systematic review and Bayesian diagnostic test accuracy network meta-analysis, PubMed, Embase, Web of Science Core Collection, and the Cochrane Library were searched from inception to March 29, 2026. Eligible studies enrolled adults with suspected or confirmed PH, evaluated imaging or imaging-derived AI models, used RHC as the main reference standard, and provided sufficient patient-level data to reconstruct 2×2 diagnostic tables. Eight prespecified nodes were compared: conventional and AI-enhanced echocardiography, computed tomography/computed tomography pulmonary angiography (CT/CTPA), conventional and AI-enhanced cardiovascular magnetic resonance (CMR), chest radiography-AI, and multimodal AI fusion. A Bayesian diagnostic test accuracy network meta-analysis estimated sensitivity, specificity, and diagnostic odds ratios, with ranking, inconsistency, meta-regression, publication-bias, and sensitivity analyses. Risk of bias was assessed with PROBAST+AI, and certainty of evidence with GRADE. The protocol was registered in PROSPERO (CRD420261431299).

Results:

Thirteen retrospective studies contributed 27 validation datasets; 22 used internal validation and 5 used external validation. CMR-AI had the highest pooled sensitivity (0.92; 95% credible interval [CrI] 0.84-0.97) but lower specificity (0.59; 95% CrI 0.31-0.82). Conventional echocardiography had the highest pooled specificity (0.92; 95% CrI 0.84-0.97) and diagnostic odds ratio (30.08; 95% CrI 10.77-66.29). CT/CTPA-AI showed higher sensitivity than conventional echocardiography (absolute difference 0.14; 95% CrI 0.00-0.29) and a higher relative diagnostic odds ratio than conventional CT/CTPA (2.21; 95% CrI 1.03-4.82). No modality simultaneously maximized sensitivity and specificity. Sensitivity analyses generally preserved the main trade-off pattern, although estimates for sparse nodes, particularly chest radiography-AI, were unstable. Meta-regression did not identify significant moderators. Evidence certainty ranged from high to very low, with very low certainty for multimodal AI fusion and concerns related to retrospective designs and frequent internal validation. Positive and negative posttest probabilities varied across nodes and assumed prevalence scenarios from 5% to 50%.

Conclusions:

AI-enhanced imaging may have a complementary role rather than replace expert interpretation or RHC. CMR-AI may support sensitive rule-out assessment, conventional echocardiography may support rule-in assessment, and CT/CTPA-AI may provide opportunistic decision support when cross-sectional imaging is already available. These findings are hypothesis-generating; prospective, multicenter, externally validated studies with standardized thresholds, calibration, and clinical-impact evaluation are needed before the results can guide routine diagnostic pathways.