Prospective study of machine learning for identification of high-risk COVID-19 patients

Abstract

The Coronavirus Disease 2019 (COVID-19) pandemic constituted a public health crisis with a devastating effect in terms of its death toll and effects on the world economy. Notably, machine learning methods have played a pivotal role in devising novel technological solutions designed to tackle challenges brought forth by this pandemic. In particular, tools for the rapid identification of high-risk COVID-19 patients have been developed to aid in the effective allocation of hospital resources and for containing the spread of the virus. A comprehensive validation of such intelligent technological approaches is needed to ascertain their clinical utility; importantly, it may help develop future strategies for efficient patient classification to be used in future viral outbreaks. Here we present a prospective study to evaluate the performance of state-of-the-art machine-learning models proposed in PloS one 16, e0257234 (2021), which we developed for the identification of high-risk COVID-19 patients across four identified clinical stages. The model relies on artificial neural networks trained with historical patient data from Mexico. To assess their predictive capabilities across the six, registered, epidemiological waves of COVID-19 infection in Mexico, we measure the accuracy within each wave without retraining the neural networks. We then compare their performance against neural networks trained with cumulative historical data up to the end of each wave. Our findings indicate that models trained using early historical data exhibit strong predictive capabilities, which allows us to accurately identify high-risk patients in subsequent epidemiological waves—under clearly varying vaccination, prevalent viral strain, and medical treatment conditions. These results show that artificial intelligence-based methods for patient classification can be robust throughout an extended period characterized by constantly evolving conditions, and represent a potentially powerful tool for tackling future pandemic events, particularly for clinical outcome prediction of individual patients.