Background: Discrimination between pre-capillary pulmonary hypertension (PH) and post-capillary PH remains challenging using current non-invasive diagnostic modalities (including echocardiography, computed tomography and magnetic resonance imaging). As a result, invasive testing with right heart catheterization (RHC) is recommended. We hypothesize that if more sensitive and specific models could be created using artificial intelligence, more confirmatory results could be obtained non-invasively.

Methods: Our study included 100 patients with a confirmed diagnosis of PH by RHC at Mayo Clinic, FL. Relevant clinical, demographic, laboratory and echocardiographic data were collected. To induce local, spatial patterns in the data for application of computer vision techniques, specifically convolutional neural networks, data columns were purposefully ordered with hierarchical clustering. All data was then scaled to be on the [0,1] interval prior to modeling. Despite the sample size being limited, 60 patients were randomly assigned to a training set, and the remaining 40 to a test set. Architecture optimization was performed using a grid search on the training data. Model performance was estimated on the test data after final selection of a model architecture, and compared to a previously developed logistic regression model. Model diagnostic capabilities were measured using the percentage of correct classifications out of all classifications (accuracy) and Area Under the Curve (AUC). All tests were two-sided and performed at an alpha level of 0.05. 

Results: CNN modeling was able to discriminate patient groups with 83% accuracy and an AUC of 0.86. When compared to a traditional logistic regression model, which identified a novel echocardiographic parameter that was highly predictive of hemodynamic PH type (AUC 0.89), CNN was non-inferior in its diagnostic capabilities to logistic regression (P=0.47). Despite the small training sample size, the predicted probabilities obtained using the CNN were well aligned with that of the logistic regression model (r=0.89, p<0.001).

Conclusions: Machine learning algorithms even on small datasets are feasible in predicting diagnostic probabilities in PH. This exploratory study highlights the potential for wider application of computer vision methods in health care.

08 - Configuring Non-Image-Based Clinical Data to Support Application of Computer Vision Techniques:  Application to Non-invasive Discrimination of PHTN