MultiCOVID: a multi modal deep learning approach for COVID-19 diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• The COVID-19 pandemic overwhelmed healthcare systems globally.
• Chest X-rays (CXR) and blood tests show predictive value for COVID-19 diagnosis.
• Accurate and rapid diagnosis is crucial for patient management.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning algorithm, MultiCOVID.
• To discriminate between COVID-19, heart failure, non-COVID pneumonia, and healthy patients.
• To improve diagnostic accuracy and speed compared to traditional methods.

Main Methods:

• Retrospective analysis of CXR and blood test data from 6123 patients (January 2017 - May 2020).
• Development of multimodal prediction models using open-source deep learning (DL) algorithms.
• Performance comparison of MultiCOVID against five experienced thoracic radiologists on 300 test images.

Main Results:

• The MultiCOVID algorithm achieved an overall accuracy of 84% with a mean AUC of 0.92 on the test set.
• For 300 random test images, MultiCOVID's accuracy (69.6%) was significantly higher than radiologists (43.7-58.7%) and consensus (59.3%).
• The algorithm demonstrated superior discrimination among the four patient groups.

Conclusions:

• A multimodal deep learning algorithm, MultiCOVID, effectively integrates CXR and blood test data.
• MultiCOVID significantly enhances diagnostic performance for COVID-19 and other respiratory conditions.
• This AI tool offers a promising approach to accelerate and improve differential diagnosis in clinical settings.