Discussion

We calculated duration of stay, etiological, and risk estimates from a large data set of hospitalized COVID-19 patients based on multistate methodology avoiding many pitfalls. The strength of multistate models is highlighted in their divergence from their traditional, yet biased alternatives. A naive (ie, not accounting for competing risks) Kaplan-Meier estimate would have shown a 30-day survival probability of 72% based on the full cohort. This finding contrasts with a survival probability of 91% from the multistate model: a massive 19% difference.

Although median length-of-stay estimates have been provided in many publications, ^{Reference Veneti, Bøås and Kristoffersen6} these estimates are invalid when there is a high degree of censoring. Notably, a median length of hospital stay for severely ill patients of 13 days could be determined by ignoring the time dependency of patients transitioning into severe illness. This naive estimate is based on knowing at admission which patients will move into the severe state (an impossible proposition) and does not give information on when or for how long they will remain in this state. The multistate result of an expected 3.5 days of severe illness (95% CI, 3.1–3.9) conditioned on patients being admitted as nonsevere on day 0 is not grounded in knowing the future and is far more informative. ‘Severe’ and ‘nonsevere’ illness have starkly different consequences in terms of costs and resources. In further contrast to traditional methods, the multistate analysis can be extended to condition on different days and states (Supplementary Material 4).

Improper treatment of current cases (by removing them from the data set) would have resulted in an adjusted odds ratio of 0.81 (95% CI, 0.59–1.14) to be discharged alive for patients in the SARS-CoV-2 ο (omicron) phase using logistic regression. This biased point estimate implies that patients in the SARS-CoV-2 ο (omicron) phase had a lower odds of being discharged alive, contrary to the properly calculated hazard ratio 1.56 (95% CI, 1.39–1.76) using Cox regression. These patients, in fact, had a higher discharge rate.

The interplay among the duration estimates, plots, and regression results is illustrative of why all are needed to attain a full account. The expected lengths of stay in the initial pandemic phase were longer than for the 3 subsequent phases. Analyzing a stacked probability plot, we observed a higher predicted death risk during the SARS-CoV-2 α (alpha) period. The influence of death would imply a shorter length of stay. However, we did not observe the higher death risk in the SARS-CoV-2 δ (delta) and SARS-CoV-2 ο (omicron) periods. In the last 2 phases, the opposing force of discharge alive (ie, quantified in the significantly higher hazard ratios) resulted in shorter hospital stays. Moreover, determining whether discharge alive (clinical improvement) or death (clinical failure) causes shorter stays is crucial for treatment comparisons. ^{Reference Martinuka, von Cube and Wolkewitz7}

It is a strength that the proposed methods can be conducted in real time, thus giving an immediate impression of changing circumstances. The use of EHRs is also advantageous in that extraction could be standardized to allow data conglomeration and meta-analyses. Although the positive-result time window facilitates this standardization, it is likely that a good proportion of patients were admitted for reasons other than COVID-19. ^{Reference Klann, Strasser and Hutch8} Further bias may have been introduced through EHR coding limitations at the outset and varying testing guidelines throughout the studied periods. The analysis was also limited by the absence of adjustment for other potential confounders such as issues of delayed care, specific treatments, or prior infection. Moreover, caution must be taken in extrapolating to other geographic regions where “severe” may have different interpretations. Lastly, calendar periods served as surrogate variables for the predominant variants, and claims have not been made regarding their intrinsic disease severity.

Alternative multistate methods to those presented might also be explored. Parametric models ^{Reference Jackson, Tom and Kirwan9} can extrapolate outside the time frame in which the data were collected, overcoming a shortcoming of nonparametric estimators. Despite not predicting the progression of patients as demonstrated here, length-of-stay estimates conditioned on the final pathway of patients can be used for resource planning. ^{Reference Keogh, Diaz-Ordaz and Jewell10}

Without ignoring the seriousness of the COVID-19 pandemic, we should learn as much as we can from this crisis. As demonstrated here, multistate methodology should be used in the analyses of future patients, whether for COVID-19 or subsequent pandemics.