Development, evaluation and validation of machine learning algorithms to detect atypical and asymptomatic presentations of Covid-19 in hospital practice.

Background: Diagnostic methods for Covid-19 have improved, both in speed and availability. Because of atypical and asymptomatic carriage of the virus and nosocomial spread within institutions, timely diagnosis remains a challenge. Machine learning models trained on blood test results have shown promise in identifying cases of Covid-19.
Aims: To train and validate a machine learning model capable of differentiating Covid-19 positive from negative patients using routine blood tests and assess the model's accuracy against atypical and asymptomatic presentations.
Design and Methods: We conducted a retrospective analysis of medical admissions to our institution during March and April 2020. Participants were categorized into Covid-19 positive or negative groups based on clinical, radiological features or nasopharyngeal swab. A machine learning model was trained on laboratory parameters and validated for accuracy, sensitivity and specificity and externally validated at an unconnected establishment.
Results: An Ensemble Bagged Tree model was trained on data collected from 405 patients (212 Covid-19 positive) producing an accuracy of 81.79% (95% confidence interval (CI) 77.53-85.55%), the sensitivity of 85.85% (CI 80.42-90.24%) and specificity of 76.65% (CI 69.49-82.84%). Accuracy was preserved for atypical and asymptomatic subgroups. Using an external data set for 226 patients (141 Covid-19 positive) accuracy of 76.82% (CI 70.87-82.08%), sensitivity of 78.38% (CI 70.87-84.72%) and specificity of 74.12% (CI 63.48-83.01%) was achieved.
Conclusion: A machine learning model using routine laboratory parameters can detect atypical and asymptomatic presentations of Covid-19 and might be an adjunct to existing screening measures.
(© The Author(s) 2021. Published by Oxford University Press on behalf of the Association of Physicians. All rights reserved. For permissions, please email: journals.permissions@oup.com.)