Using machine learning to improve our understanding of COVID-19 infection in children

Abstract

Purpose: Children are at elevated risk for COVID-19 (SARS-CoV-2) infection due to their social behaviors. The purpose of this study was to determine if usage of radiological chest X-rays impressions can help predict whether a young adult has COVID-19 infection or not.

Methods: A total of 2572 chest impressions from 721 individuals under the age of 18 years were considered for this study. An ensemble learning method, Random Forest Classifier (RFC), was used for classification of patients suffering from infection.

Results: Five RFC models were implemented with incremental features and the best model achieved an F1-score of 0.79 with Area Under the ROC curve as 0.85 using all input features. Hyper parameter tuning and cross validation was performed using grid search cross validation and SHAP model was used to determine feature importance. The radiological features such as pneumonia, small airways disease, and atelectasis (confounded with catheter) were found to be highly associated with predicting the status of COVID-19 infection.

Conclusions: In this sample, radiological X-ray films can predict the status of COVID-19 infection with good accuracy. The multivariate model including symptoms presented around the time of COVID-19 test yielded good prediction score.

Fig 1. Flow chart used to divide Chest X-Ray impressions into three categories.

The scans were preprocessed and then given as an input for categorization. These categories include Normal, Stable, and New findings.

Fig 2. Frequencies of each feature in COVID-19 positive and negative patients.

Fig 3. Random forest classifier performance for identifying COVID-19 infection.

The ROC curves shown above identifies the ability of all ML models to classify pediatric patients with COVID-19 infection.

Fig 4. SHAP values for most important model features in model 5 shown in decreasing order of their importance along the Y-axis.

Features in the upper case indicate RoS data, the lower case without an underscore represents features obtained from CXRi, and some additional demographic features. The top 20 features for the Random Forest classifier using a total of 76 features are shown using model 4. Each dot on the X-axis represents the importance value of the corresponding feature for each patient. The location of each dot indicates whether the feature is positively or negatively associated with the output. The color of each dot indicates whether the value is high (shown in red) or the value is low (indicated in blue).

Fig 5

SHAP values for two patients with absence (a) and presence (b) of COVID-19 infection. The SHAP values above indicate the impact of a particular feature with a certain value in comparison to the prediction made if the feature took some baseline value. As observed in (a), the absence of pneumonia, atelectasis, and small airways disease indicates the absence of COVID-19 infection and the presence of these features in (b) indicates the presence of COVID-19 infection.