Machine-learning tool for classifying pulmonary hypertension via expert reader-provided CT features: An educational...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Diseases

Background:

• Pulmonary hypertension (PH) is a complex condition categorized into five groups by ESC/ERS guidelines.
• Chest contrast-enhanced computed tomography (CECT) is vital for non-invasive PH assessment.
• Accurate classification of PH is essential for appropriate patient management.

Purpose of the Study:

• To develop a machine learning (ML)-based educational resource for classifying PH cases using CECT.
• To categorize PH according to the established ESC/ERS groups (I-V).
• To create a tool that aids in the non-invasive assessment of PH.

Main Methods:

• Retrospective analysis of 172 PH patients' CECT scans from two hospitals.
• Independent review of CECTs by three radiologists, extracting 13 features and assigning PH group likelihood scores.
• Development and testing of various ML algorithms (e.g., Naïve-Bayes) using Weka software, with performance evaluated by accuracy, AUROC, and F1-score.

Main Results:

• The Naïve-Bayes algorithm achieved 0.72 accuracy, 0.84 AUROC, and 0.72 F1-score after excluding three group V patients.
• Specific performance metrics varied across PH groups I-IV, with AUROC values generally high (0.78-0.87).
• The study demonstrated the feasibility of using ML for PH classification from CECT data.

Conclusions:

• This is the first study to create an ML tool for PH classification via chest CECT.
• The developed resource has the potential to train radiologists in PH classification, supporting multidisciplinary decision-making.
• Further improvements, including a larger dataset, are necessary to enhance the tool's performance metrics.