Comparison of deep learning with regression analysis in creating predictive models for SARS-CoV-2 outcomes

Area of Science:

• Medical Informatics
• Computational Biology
• Epidemiology

Background:

• Accurate prediction of patient outcomes in Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is vital for effective patient management and healthcare resource allocation.
• Existing prognostic models for SARS-CoV-2 have shown susceptibility to bias.
• Developing robust predictive models is essential for improving clinical decision-making.

Purpose of the Study:

• To develop and compare two distinct predictive models for in-hospital mortality in patients with SARS-CoV-2.
• To evaluate the performance of a Cox regression model and an artificial neural network (ANN) model.
• To assess the accuracy, discrimination, and calibration of the developed prognostic models.

Main Methods:

• Utilized electronic health record data from 398 laboratory-confirmed SARS-CoV-2 patients admitted to a London teaching hospital between March 1 and April 24, 2020.
• Developed two predictive models: a Cox regression model and an artificial neural network (ANN).
• Assessed model performance using validation, discrimination (Area Under the Receiver Operator Curve - AUROC), and calibration (Brier score).

Main Results:

• Both Cox regression and ANN models demonstrated high accuracy (83.8% and 90.0%, respectively).
• The ANN model achieved a significantly greater AUROC (92.6%) compared to the Cox regression model (86.9%), indicating superior discrimination (p=0.0136).
• Both models exhibited acceptable calibration, with Brier scores of 0.13 for the Cox model and 0.11 for the ANN.

Conclusions:

• An artificial neural network (ANN) model for predicting SARS-CoV-2 mortality is non-inferior to a Cox regression model.
• The ANN model offers potential for further development and adaptation as new data becomes available.
• Accurate prognostic models, particularly those leveraging deep learning for complex datasets, can significantly benefit patient care and healthcare systems.