Development of a machine-learning-based CT educational tool for grouping pulmonary hypertension: A proof-of-concept study.

Purpose/Objectives: Pulmonary hypertension (PH) is a multifaceted disease with different etiologies and clinical presentations. The European Society of Cardiology (ESC) and the European Respiratory Society (ERS) joint guidelines for PH classification include recommendations on the use of chest contrastenhanced computed tomography (CECT) [1]. While CECT plays a pivotal role in this setting [2] [3], there is a need to train nondedicated radiologists to interpret it accurately for adequate classification and management (Figure 1). We aimed to develop a machine learning-driven educational tool for categorizing PH via CECT. Methods and Materials: The study retrospectively included 141 patients diagnosed with PH who underwent chest multidetector CECT between 2014 and 2023 at Center 1 (94/141) or Center 2 (47/141) University Hospital. Three readers experienced in thoracic imaging independently reviewed 47 randomly selected exams each. Readers reported 13 CECT features, including lung conditions (emphysema, airway disease, lung fibrosis, centrilobular ground-glass nodules, and mosaic pattern), heart abnormalities (including left atrium [LA] enlargement, coronary calcifications, valvular calcifications, and congenital heart diseases), signs of chronic thromboembolism, esophageal dilatation, lymph node enlargement, and portal hypertension (Figure 2). Related classification of PH cases into groups 1-5 as per the ESC/ERS guidelines [1] constituted the ground truth. We exported suitable features to an attribute-relation file format (ARFF). The ARFF file was then uploaded to the machine-learning Weka software (Waikato, New Zealand) [4]. Several machine-learning algorithms inside Weka were experimented with a 10-fold validation process. Results: We removed the group 5 PH cases due to scarce numerosity (3/141). NaiveBayes was the algorithm with the best performance, correctly classifying 102 out of 138 cases. The algorithm led to a substantial agreement between the output and ground truth (Cohen kappa = 0.644), with an accuracy of 0.739. Group 1 PH was the class with the worst performance, with 20/29 (69%) cases correctly attributed. Table 1 shows the confusion matrix. Conclusion: The best machine-learning educational tool we tested showed reasonable accuracy in classifying PH, thus having the potential for clinical use and education of less experienced readers.