Symptom-Based Predictive Model of COVID-19 Disease in Children

Abstract

Background: Testing for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection is neither always accessible nor easy to perform in children. We aimed to propose a machine learning model to assess the need for a SARS-CoV-2 test in children (<16 years old), depending on their clinical symptoms.

Methods: Epidemiological and clinical data were obtained from the REDCap^® registry. Overall, 4434 SARS-CoV-2 tests were performed in symptomatic children between 1 November 2020 and 31 March 2021, 784 were positive (17.68%). We pre-processed the data to be suitable for a machine learning (ML) algorithm, balancing the positive-negative rate and preparing subsets of data by age. We trained several models and chose those with the best performance for each subset.

Results: The use of ML demonstrated an AUROC of 0.65 to predict a COVID-19 diagnosis in children. The absence of high-grade fever was the major predictor of COVID-19 in younger children, whereas loss of taste or smell was the most determinant symptom in older children.

Conclusions: Although the accuracy of the models was lower than expected, they can be used to provide a diagnosis when epidemiological data on the risk of exposure to COVID-19 is unknown.

Keywords: COVID-19; SARS-CoV-2; deep learning; epidemiology; machine learning; microbiology; paediatrics.

Figure 1

Modelling and data pipeline: a classifier f^s and its quality metrics is obtained for each dataset X^s.

Figure 2

Impact of each variable on the model output for the general model. The features are organised top-down, by decreasing overall importance. For each feature, the SHAP value of each test observation is shown as a point. Each symptom is present if the colour of the point is red and absent if it is blue. The more to the right the points are, the more the output is associated with a SARS-CoV-2 infection.

Figure 3

(A) Absolute mean impact and (B) absolute maximum impact of each variable for the general model.

Figure 4

(A) Impact of each variable on the model output for the model for children 0 to 5 years old and (B) for children 6 to 15 years old. The features are organised top-down, by decreasing overall importance. For each feature, the SHAP value of each test observation is shown as a point. Each symptom is present if the colour of the point is red and absent if it is blue. The more to the right the points are, the more the output is associated with a SARS-CoV-2 infection.

Figure 4

(A) Impact of each variable on the model output for the model for children 0 to 5 years old and (B) for children 6 to 15 years old. The features are organised top-down, by decreasing overall importance. For each feature, the SHAP value of each test observation is shown as a point. Each symptom is present if the colour of the point is red and absent if it is blue. The more to the right the points are, the more the output is associated with a SARS-CoV-2 infection.