What is the optimal approach to analyze ventilator-free days? A simulation study

Position du problème et objectif(s) de l’étude

Ventilator-free days (VFDs) are a composite outcome in critical care research, reflecting both survival and mechanical ventilation duration [1]. However, analysis methods for VFDs are inconsistent, with some focusing on count-based approaches and others on time-to-event outcomes, while other approaches such as the multistate model and the win ratio have emerged [2]. We aimed to evaluate various statistical models through simulations to identify the optimal approach for analyzing VFDs.

Matériel et méthodes

First, 16 datasets of 300 individuals were simulated, comparing a control group to an intervention with varying survival rates and ventilation durations. A total of 3,000 independent replications were performed to estimate the statistical power, with an accuracy of at least 0.01, and the Type I error rate. A clinical trial dataset (LIVE study, NCT02149589) [3] was then used to apply the same statistical models to analyze VFDs. Eleven statistical methods were evaluated, including count-based (the Mann-Whitney test, zero-inflated negative binomial model, negative binomial hurdle model, zero-inflated Poisson model, and Poisson hurdle model), time-to-event approaches (the log-rank test, the Gray test, the cause-specific Cox model, the Fine-Gray model, and the multistate model), and the win-ratio. Additionally, sensitivity analyses were conducted.

Résultats & Discussion

Most statistical methods effectively controlled Type I error rate, except for the zero-inflated and hurdle Poisson/negative binomial count submodels, as well as the cause-specific Cox regression model for death (Figure 1). The power to detect survival benefit and ventilation duration effects varied, with time-to-event approaches, mainly the multistate model, and the win ratio generally performing best. Similar results were observed in sensitivity analyses. In the LIVE dataset, all methods with Type I error control showed a significant association between VFDs and randomization groups. Graphical representations of the different types of statistical approaches (i.e., count-based approach of VFDs, time-to-event analysis with competitive risks, or the multistate model) are provided in Figure 2. Our study focused on VFDs while there are many other outcomes based on event-free days being used in critical care research, thus reinforcing the importance of the current results. Similarly, while our simulated trials focused solely on invasive mechanical ventilation, our findings could be extrapolated to include other methods of ventilatory support, such as non-invasive ventilation.

Conclusion

In this study based on simulations to evaluate multiple models that could be used to analyze VFDs, count-based approaches had low performance, while the multistate model, the Mann-Whitney test, and the win ratio were the best models for statistical power and control of Type I error. Given the lack of effect size provided by the Mann-Whitney test and potential difficulties in interpreting the win ratio, the multistate model should be the preferred option to use to analyze VFDs.

Auteurs

Laurent RENARD TRICHÉ (1) , Matthieu JABAUDON (2), Nicolas MOLINARI (3), Jean-Michel CONSTANTIN (4), Bruno PEREIRA (5), Sylvie CHEVRET (6) - (1)Department Of Perioperative Medicine, Chu Clermont-Ferrand, Clermont-Ferrand, France ; Igred, Inserm, Cnrs, Université Clermont Auvergne, Clermont-Ferrand, France ; Ecstrra Team, Irsl, Inserm Umr1342, Université Paris Cité, Paris, France., Clermont-Ferrand, France, (2)Department Of Perioperative Medicine, Chu Clermont-Ferrand, Clermont-Ferrand, France ; Igred, Inserm, Cnrs, Université Clermont Auvergne, Clermont-Ferrand, France., Clermont-Ferrand, France, (3)Idesp, Inserm, Premedical Inria, Univ Montpellier, Chu Montpellier, Montpellier, France., Montpellier, France, (4)Sorbonne University, Grc 29, Ap-Hp, Dmu Dream, Department Of Anesthesiology And Critical Care, Pitié-Salpêtrière Hospital, Paris, France., Paris, France, (5)Biostatistics Unit, Department Of Clinical Research, And Innovation (drci), Chu Clermont-Ferrand, Clermont-Ferrand, France., Clermont-Ferrand, France, (6)Ecstrra Team, Irsl, Inserm Umr1342, Université Paris Cité, Paris, France ; Department Of Biostatistics, Hôpital Saint-Louis, Ap-Hp, Paris, France., Paris, France

Orateur(s)

Laurent RENARD TRICHÉ (Clermont-Ferrand)