Deep learning model outperforms echocardiography in identifying pulmonary hypertension

Key takeaways:

• The model combines demographic, chest X-ray, electrocardiogram and transthoracic echocardiography data.
• The model’s accuracy was validated across various clinical severities and demographics.

BOSTON — A multimodal fusion model outperformed transthoracic echocardiography in identifying pulmonary hypertension, according to a study presented at the CHEST Annual Meeting.

This model could reduce the need for invasive procedures and facilitate personalized treatment, Zhihua Huang, MD, PhD, Fuwai Hospital, CAMS and PUMC, Beijing, and colleagues wrote.

Transthoracic echocardiography (TTE) and other diagnostic tools require invasive right heart catheterization to confirm pulmonary hypertension. However, the Multimodal Fusion Model for Pulmonary Hypertension (MMF-PH) uses data from demographics, chest X-rays, electrocardiograms, and TTE for diagnosis.

The researchers began their study with data from 4,576 patients who had suspected pulmonary hypertension, including 2,451 whose pulmonary hypertension was confirmed, in addition to a control group.

During training, validation and testing phases, MMF-PH also tapped a prospective dataset of 477 patients and an external dataset from two other medical facilities as part of its deep learning.

Next, the researchers compared the efficacy of MMF-PH against TTE and another derived model in delineating hemodynamic subtypes of pulmonary hypertension.

The researchers reported that MMF-PH markedly surpassed TTE in detecting pulmonary hypertension. The model’s retrospective accuracy was 96.2%, and its areas under the receiver operating characteristic (AUROC) curve was 0.994.

Prospective analysis of MMF-PH indicated a 0.969 AUROC and enhanced specificity of 96.9% and sensitivity of 95.4% compared with TTE, along with an F1 score of 0.961 and significant improvements in predictive values, the researchers said.

The retrospective analysis further indicated specificities of 70% for MMF-PH and 53.3% for TTE at equal sensitivity rates. In the prospective analysis, specificities included 89.1% for MMF-PH and 65.2% for TTE.

Across various demographics and clinical severities, the researchers continued, the accuracy of MMF-PH was consistently validated in subgroup analyses.

External validation yielded a 92% sensitivity and an 87.9% F1 score, the researchers added, reinforcing the reliability of MMF-PH.

Further, the 96.9% sensitivity for non-pulmonary hypertension and 82.7% sensitivity for pre-capillary pulmonary hypertension in the retrospective cohort showed that MMF-PH effectively distinguished between pulmonary hypertension subtypes, the researchers said.

These findings indicated that MMF-PH significantly advances the detection and subtyping of pulmonary hypertension, the researchers concluded, with accuracy that is superior to traditional TTE.

The researchers also said the model exhibited high diagnostic performance with extensive applicability across a range of patient groups and clinical environments, indicating its potential for broad clinical adoption.

These characteristics may enable the early and accurate identification of pulmonary hypertension, the researchers said, which would reduce the need for right heart catheterization and other invasive diagnostic procedures.

MMF-PH also may facilitate personalized treatment approaches, the researchers continued, by accurately classifying subtypes of pulmonary hypertension to enhance patient management, optimize treatment, improve outcomes and reduce costs.