Abstract
Estimating the lengths-of-stay (LoS) of hospitalised COVID-19 patients is key for predicting the hospital beds' demand and planning mitigation strategies, as overwhelming the healthcare systems has critical consequences for disease mortality. However, accurately mapping the time-to-event of hospital outcomes, such as the LoS in the intensive care unit (ICU), requires understanding patient trajectories while adjusting for covariates and observation bias, such as incomplete data. Standard methods, such as the Kaplan-Meier estimator, require prior assumptions that are untenable given current knowledge. Using real-time surveillance data from the first weeks of the COVID-19 epidemic in Galicia (Spain), we aimed to model the time-to-event and event probabilities of patients' hospitalised, without parametric priors and adjusting for individual covariates. We applied a non-parametric mixture cure model and compared its performance in estimating hospital ward (HW)/ICU LoS to the performances of commonly used methods to estimate survival. We showed that the proposed model outperformed standard approaches, providing more accurate ICU and HW LoS estimates. Finally, we applied our model estimates to simulate COVID-19 hospital demand using a Monte Carlo algorithm. We provided evidence that adjusting for sex, generally overlooked in prediction models, together with age is key for accurately forecasting HW and ICU occupancy, as well as discharge or death outcomes.
Highlights
As of January 2021, SARS-CoV-2 transmission continues to increase in most countries worldwide [1], and in those countries where control has been achieved, resurgences are expected [2] before effective vaccines are widely available
When an event happens for all patients (a.s. ‘leave the hospital’, when all status on discharge gathered as a composite outcome), KM is not biased and coincides with the NP-mixture cure models (MCMs), both of them represented with the one single line in Figure 1 and Supplementary Figure S2
When the final outcome is experienced by only a proportion of patients (‘admission to intensive care unit (ICU)’, ‘death’, ‘discharge’), the KM overestimates the time-to-event showing longer LoS than the non-parametric MCM (NP-MCM)
Summary
As of January 2021, SARS-CoV-2 transmission continues to increase in most countries worldwide [1], and in those countries where control has been achieved, resurgences are expected [2] before effective vaccines are widely available. Understanding and predicting inpatient lengths-of-stay (LoS) and critical-care demand remain some of the major components of outbreak monitoring for decision-making and contingency planning. Predicting hospital demand entails estimating a patient’s LoS and the probability of hospital outcomes such as requiring admission to the intensive care unit (ICU). Estimation of these variables is challenging as it requires investigating the patients’ trajectories, and it must account for complexities in the data. The LoS of some inpatients may be censored because the study ends before the patient leaves the hospital facility. The LoS of COVID-19 patients has been studied using parametric models [4], semi-parametric methods [5] and non-parametric estimators [3, 6]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
