Predicting COVID-19 incidences from patients? Viral load using deep-learning

, , , , , et

référence BibTeX

The transmission of the contagious Coronavirus disease (COVID-19) is highly dependent on individual viral dynamics. Reverse-transcription quantitative polymerase chain reaction (RT-qPCR) tests used for diagnosing COVID-19 provide a semi-quantitative measurement of viral load within the infected host in the form of a Cycle treshold (Ct) value. We solicited Ct values sampled from a cross-sectional patient cohort at Rafik Hariri University Hospital to now-cast COVID-19 incidences in Lebanon. Our patient cohort of 9531 patients, retrieved from a single representative cross-sectional virology test center in Lebanon, revealed that when the mean Ct value of a daily sample of patients is low, an increase in positive COVID-19 case counts is observed in Lebanon. A subset of the data was used to train several machine learning models and tune their hyperparameters with respect to the validation error. Unseen data unused during model development is used to report the models' test error. Support vector machine regression performed well on the unseen data and was able to provide predictions for the positive case counts in Lebanon over the span of 7 days. The models are a first attempt at capturing the interaction between cross-sectional Ct values and the pandemic trajectory including temporal effects that arise from population dynamics. The model has potential applications for predicting COVID-19 incidences in other countries and in assessing post-vaccination policies. Apart from emphasizing patient responsibility in adopting early testing practices, this study proposed and validated viral load measurement as a relevant input that can enhance the predictive accuracy of viral disease now-casting models.

, 12 pages

Ce cahier a été révisé en mai 2022

Axe de recherche

Application de recherche


G2148R.pdf (1,6 Mo)