Machine-learning tool for classifying pulmonary hypertension via expert reader-provided CT features: An educational resource for non-dedicated radiologists

Abstract

Purpose

Pulmonary hypertension (PH) is a complex disease classified into five groups (I-V) by the European Society of Cardiology/European Respiratory Society (ESC/ERS) guidelines. Chest contrast-enhanced computed tomography (CECT) is crucial in the non-invasive PH assessment. This study aimed to develop a machine learning (ML)-based educational resource for classifying PH cases via CECT according to ESC/ERS groups.

Methods

We retrospectively included 172 PH patients who underwent CECT at two University Hospitals (Udine and Brescia). Three chest-devoted radiologists independently reviewed the CECTs, reporting on 13 features, including lung conditions, heart abnormalities, chronic thromboembolism, and mediastinal findings. Readers assigned the features as absent/present except for the left atrium (LA) anteroposterior diameter (measured in millimeters) and classified PH cases I-V with likelihood scores (1–100 %) for each group. The majority decisions for features and average LA diameter were used as ML inputs. The highest average likelihood scores determined group assignments, serving as ground truth. Various ML algorithms were tested using the Weka software and evaluated by accuracy, area under the ROC curve (AUROC), and F1-score.

Results

After excluding three group V patients to avoid imbalance, the Naïve-Bayes algorithm showed 0.72 accuracy, 0.84 AUROC, and 0.72 F1-score. Accuracy values for group I-IV were 0.75, 0.78, 0.51, 0.79; AUROC values were 0.78, 0.84, 0.86, 0.87; F1-scores were 0.63, 0.79, 0.61, 0.84, respectively.

Conclusions

This study is the first to develop an ML-driven tool for classifying PH via chest CECT. While performance metrics require improvement, including the need for a larger sample size, the resource can potentially train non-dedicated radiologists in PH classification, supporting multidisciplinary reasoning.