Machine learning for the prediction of pathologic pneumatosis intestinalis

Area of Science:

• Radiology
• Artificial Intelligence
• Gastroenterology

Background:

• Pneumatosis intestinalis is a radiographic finding with a wide range of clinical implications, from benign to life-threatening.
• Current management lacks a validated tool to guide surgical decisions despite the commonality of radiographic pneumatosis intestinalis.

Purpose of the Study:

• To develop and evaluate machine learning models for distinguishing benign from pathologic pneumatosis intestinalis.
• To determine which patients with pneumatosis intestinalis would benefit from surgical intervention.

Main Methods:

• Developed three machine learning models: an imaging model (radiomic features from CT scans), a clinical model (clinical variables), and a combination model (imaging + clinical variables).
• Retrospectively analyzed 300 cases of pneumatosis intestinalis from a single institution.

Main Results:

• The combination model significantly outperformed individual imaging and clinical models for both tasks.
• Achieved an area under the curve (AUC) of 0.91 for distinguishing benign from pathologic cases and 0.84 for predicting surgical benefit.
• The clinical model had an AUC of 0.87 for task 1 and 0.76 for task 2; the imaging model had an AUC of 0.72 for task 1 and 0.68 for task 2.

Conclusions:

• Combined radiographic and clinical features effectively identify pathologic pneumatosis intestinalis.
• This approach aids in selecting patients who require surgery, potentially improving the surgical decision-making process.