A Hybrid Decision Tree and Deep Learning Approach Combining Medical Imaging and Electronic Medical Records to Predict...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Clinical Decision Support Systems

Background:

• Early prediction of invasive mechanical ventilation (IMV) is crucial for managing COVID-19 patients and optimizing resource allocation.
• Machine learning (ML) models offer a systematic approach to prognostic assessment for patients at risk of respiratory failure.
• Chest radiographs (CXRs) and electronic medical records (EMRs) are key early indicators for IMV need in COVID-19 hospitalizations.

Purpose of the Study:

• To evaluate an ML model for predicting the need for intubation within 24 hours.
• To utilize a combination of CXR and EMR data within an automated pipeline.
• To analyze historical data from 2481 hospitalizations at The Mount Sinai Hospital.

Main Methods:

• Automated preprocessing of CXRs including resizing, rescaling, normalization, and lung segmentation using U-Net.
• Training an image classifier (DenseNet) with transfer learning, 10-fold cross-validation, and grid search on augmented CXR data.
• Developing a fusion model combining the image classifier's probability score with 41 longitudinal EMR variables using a random forest algorithm.

Main Results:

• The fusion model achieved 78.9% sensitivity and 83% specificity at a 0.5 probability threshold.
• The model demonstrated a significant improvement over the image classifier alone, with an AUROC of 0.874 and AUPRC of 0.497.
• Key predictors included respiratory rate, C-reactive protein, oxygen saturation, and lactate dehydrogenase; imaging score ranked 15th in importance.

Conclusions:

• An automated deep learning image classifier, integrated with EMR data, enhances the identification of severe COVID-19 patients requiring intubation.
• This model shows potential for assisting risk assessment and optimizing clinical decision-making during critical COVID-19 care.
• Further prospective and external validation is recommended for broader clinical application.