Can adaptive heterogeneous graph neural networks differentiate PAH from PH-LHD?

Two PH groups with different treatment rules

The clinical problem is a classification fork, not a detection task. Pulmonary arterial hypertension (PAH) is a pre-capillary group, while pulmonary hypertension associated with left heart disease (PH-LHD) is Group 2 PH and may be isolated post-capillary or combined post- and pre-capillary depending on pulmonary vascular resistance. The paper notes that drugs approved for PAH are not recommended for routine treatment of PH-LHD, and that trials reported fluid-retention concerns with macitentan in combined post- and pre-capillary PH-LHD and worse outcomes with sildenafil in persistent PH after corrected valvular disease.

Definitive sorting still requires right heart catheterization (RHC) read together with imaging, pulmonary testing and history. The authors therefore frame any noninvasive model as a triage adjunct that flags who needs specialist assessment, not as a replacement for RHC. That framing matters when reading the performance numbers: the model was trained to reproduce an RHC-anchored label, not to decide whether catheterization happens.

Labels came from invasive haemodynamics

The cohort was drawn from adults evaluated at Shanghai Pulmonary Hospital between January 2015 and December 2022 who underwent RHC for suspected PH. From 3,250 screened patients, exclusions for Group 3, Group 4 and Group 5 PH, emergency catheterization or instability, missing or non-diagnostic CT, and more than 20% missing core noninvasive variables left 943 for reference-label adjudication; 38 more were excluded, giving 905 patients — 421 with PAH and 484 with PH-LHD.

A technically reliable pulmonary arterial wedge pressure (PAWP) was available in 713 of 905 patients (78.8%). For the remaining 192 (21.2%), invasively measured left ventricular end-diastolic pressure from left heart catheterization during the same hospitalization supported the left-sided filling pressure classification. Forty-five patients near the boundary (PAWP 13–15 or 16–18 mmHg, or the equivalent LVEDP bands) were adjudicated using the haemodynamic value together with left atrial size, mitral E/e' ratio and evidence of left heart disease. Mean left-sided filling pressure was 10 ± 3 mmHg in PAH and 24 ± 9 mmHg in PH-LHD.

Critically, no RHC-derived variable entered model development, training or inference. Invasive data defined the answer key and described the cohort; the model saw only CT and noninvasive variables.

Edges are learned rather than fixed

AHGNN treats each patient as a node in a heterogeneous network and learns which patients should be connected, instead of imposing one fixed similarity metric across demographics, ECG, echocardiography, pulmonary function, laboratory values and imaging. The paper's argument is that a single fixed metric forces an inappropriate topology when different variable families encode different relationships. Graph construction is made differentiable with Gumbel-Softmax reparameterization, so the similarity structure is trained jointly with the classifier.

Two further components support this. Hierarchical cross-modal attention models feature interactions at multiple levels, and dual-level self-supervised contrastive learning aims at robust representations when labeled cohorts are limited. The classifier outputs an estimated probability of PH-LHD; one minus that value is the estimated probability of PAH, an orientation used consistently throughout.

The learned clusters are descriptive, not clinical. In the representative fold with median internal-validation AUC, the model grouped variables into four types — for example age, BMI, LVEDD, RV dilation, PEF, SV1+RV6, haemoglobin and haematocrit in Type 1; sex, diabetes, hypertension, atrial fibrillation, PR interval and QRS duration in Type 2. The paper states these are model-derived clusters that should not be read as clinical classifications, and that attention weights describe model behaviour rather than causal or disease-specific biological importance.

Adaptive graphs beat fixed graphs and plain fusion

The headline comparison is against models fitted on identical outer partitions with fold-restricted preprocessing. AHGNN reached ROC-AUC 0.946 ± 0.023 and PR-AUC 0.952 ± 0.009. A fixed-graph HetNet reached 0.783, a vanilla GCN 0.702, and a clinical-only penalized multivariable logistic regression 0.778 ± 0.039 (PR-AUC 0.792 ± 0.038).

Modality-specific and fusion baselines were weaker in the same direction: TextNet 0.869 ± 0.010, ImageNet on CT alone 0.716 ± 0.011, and a multimodal MLP 0.784 ± 0.012. The authors read the drop in text-only and image-only variants as evidence that structured clinical variables and CT-derived features contributed complementary information. They also note that the near-equivalence of the clinical MLR and the multimodal MLP indicates simple feature fusion alone did not explain AHGNN's performance.

The paper is explicit that the size of these gaps should be read cautiously, because the evaluation was retrospective and single-centre and has not been reproduced in an independent cohort. The authors also tested whether age and comorbidity burden could explain the result, refitting models with age alone, atrial fibrillation alone, diabetes alone, hypertension alone, an eight-variable case-mix set, and AHGNN with case-mix variables removed; those comparisons used paired empirical 95% intervals across the 100 fold-level differences, which the paper describes as descriptive rather than formal confidence intervals or hypothesis tests.

Sensitivity and specificity trade off by threshold

PH-LHD was the positive class, and thresholds were selected only on the corresponding internal validation subset — never on outer test data. At the fixed 0.50 threshold, sensitivity was 0.867 ± 0.070 and specificity 0.867 ± 0.072, with PPV 0.888 ± 0.048 and NPV 0.858 ± 0.057.

The validation-derived Youden threshold averaged 0.504 ± 0.045 and raised both: sensitivity 0.918 ± 0.029, specificity 0.924 ± 0.041, PPV 0.934 ± 0.032, NPV 0.910 ± 0.027. A high-sensitivity triage threshold, chosen to reach at least 90% sensitivity for PH-LHD, averaged 0.456 ± 0.045 and produced sensitivity 0.930 ± 0.025 with specificity 0.791 ± 0.085.

These are different operating points answering different questions, and the paper keeps them separate. PPV and NPV also depend on the 53.5% prevalence of PH-LHD in this referral cohort, so they should not be carried unchanged to a population with a different case mix.

Calibration is the weak link

Discrimination and calibration diverge here. Pooled outer-test predictions gave a Brier score of 0.139, a calibration intercept of 0.149 and a calibration slope of 3.547. A slope well above 1 means predicted probabilities were insufficiently dispersed relative to observed outcomes: the model ranks patients well but its probability values are too compressed. The authors conclude recalibration is likely necessary before prospective use.

Decision-curve analysis showed greater net benefit for AHGNN than classifying all or no patients as PH-LHD across threshold probabilities from 0.05 to 0.80. The paper labels this exploratory, because thresholds, prevalence and probability calibration have not been validated externally, and states that internal net benefit does not establish clinical utility in an external population.

Single-centre evidence with a triage-only claim

Several boundaries sit close to the results. The cohort is retrospective and from one hospital, with an observed age range of 41–84 years and a referral case mix in which PH-LHD was 53.5% of patients. CT came from two scanner platforms — GE Revolution (n = 543) and Siemens Force (n = 362) — and scanner platform was completely associated with tube voltage and reconstructed slice thickness, so scanner-stratified performance was evaluated descriptively rather than as an isolated factor.

The paper also reports that between-group differences in age, sex, BMI and smoking indicate the model may partly learn comorbidity and referral patterns, a point the authors say they considered when interpreting variable importance. Subgroup analyses by age, sex and atrial-fibrillation status were descriptive, with no formal interaction testing.

The stated path forward is prospective external validation, recalibration and clinical safety evaluation before clinical use or any change to the role of right heart catheterization. The authors are clear that retrospective discrimination and internally estimated net benefit do not establish that any patient can safely forgo RHC.

Recalibration before any referral use

The paper's most decision-relevant result is not the 0.946 AUC but the pairing of that AUC with a calibration slope of 3.547. A model that ranks well yet emits compressed probabilities cannot be dropped into a referral pathway at a fixed cutoff without recalibration, and the threshold table shows how much the error mix moves between 0.50, the Youden point near 0.504 and the high-sensitivity point near 0.456.

Read together, the evidence supports one specific priority: take this multimodal, adaptively graphed approach into an independent cohort, recalibrate it there, and re-examine the high-sensitivity operating point that trades specificity (0.791) for sensitivity (0.930). Until that happens, the appropriate claim is internal proof of concept for adaptive graph learning in PH phenotyping, with RHC unchanged as the definitive classifier.

Complementary inputs, unproven mechanism

The ablation pattern — weaker text-only and image-only variants, and a multimodal MLP that did not match AHGNN — is consistent with the authors' explanation that structured clinical variables and CT features carry complementary information and that adaptive graph learning adds something beyond simple fusion. It is not proof of that mechanism. The paper itself says the contribution of adaptive graph learning requires confirmation in independent data, and it treats the learned variable types and attention weights as descriptions of model behaviour rather than biological findings.

That distinction sets the reader's next step. The useful question is not whether AHGNN outperformed fixed-graph and fusion baselines in this cohort — it did, by a wide margin on ROC-AUC — but whether the learned similarity structure and the probability scale survive transport to a new hospital, a different scanner mix and a different prevalence of PH-LHD.

Sources used in this answer