Deep Learning Models for Predicting Severe Progression in COVID-19-Infected Patients: Retrospective Study

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Deep Learning for Disease Prediction

Background:

• COVID-19 progression varies significantly, with some patients rapidly developing respiratory failure.
• Early identification of high-risk COVID-19 cases is crucial for timely medical intervention and improved outcomes.
• Computed Tomography (CT) imaging offers valuable insights into disease severity.

Purpose of the Study:

• To develop and validate deep learning models for the rapid identification of high-risk COVID-19 patients.
• To integrate both CT imaging and clinical data for enhanced predictive accuracy.
• To create a tool that assists clinicians in stratifying COVID-19 patient risk.

Main Methods:

• A cohort of 297 COVID-19 patients from five South Korean hospitals was analyzed.
• A mixed Artificial Convolutional Neural Network (ACNN) model was created, combining Artificial Neural Network (ANN) for clinical data and Convolutional Neural Network (CNN) for 3D CT images.
• The ACNN model was trained to classify patients into high-risk (event) or low-risk (event-free) progression groups.

Main Results:

• The mixed ACNN model demonstrated high performance in classifying risk using novel coronavirus pneumonia lesion images (93.9% accuracy, 0.916 AUC).
• Classification using lung segmentation images also yielded strong results (94.3% accuracy, 0.928 AUC).
• The model effectively differentiated between high-risk and low-risk COVID-19 progression groups.

Conclusions:

• The developed deep learning model successfully identifies high-risk COVID-19 patients by integrating imaging and clinical data.
• This predictive tool can facilitate early intervention strategies and guide aggressive therapeutic approaches.
• The study highlights the potential of AI in managing COVID-19 patient care and resource allocation.