Development and validation of multimodal deep learning algorithms for detecting pulmonary hypertension.

Summary

A multimodal fusion deep learning model (MMF-PH) trained on 2,451 right-heart-catheterized patients, with prospective (n=477) and external validation, outperformed standard echocardiography in specificity and negative predictive value for pulmonary hypertension screening. An ablation study confirmed the necessity of each module, and performance was robust across subgroups.

Key Findings

• MMF-PH outperformed standard TTE in specificity and negative predictive value across multiple test datasets.
• Training on 2,451 RHC-confirmed cases with prospective (n=477) and external validation supports generalizability.
• Ablation analysis demonstrated each model module contributes meaningfully to performance.
• Performance remained robust across diverse patient subgroups, enhancing clinical reliability.

Clinical Implications

Adoption of MMF-PH could reduce false positives, prioritize referrals for right heart catheterization, and enable earlier PH detection, especially in resource-constrained settings by augmenting TTE with AI decision support.

Limitations

• Abstract lacks granular metrics (e.g., AUROC, sensitivity) and detailed modality composition.
• Model interpretability and performance in non-participating centers require further evaluation.

Future Directions

Head-to-head implementation studies assessing clinical workflows, downstream RHC utilization, and outcomes; calibration/interpretability work; multinational validations and regulatory pathways.