Multistate Models in Critical Care: Two Steps Forward, One Step Back

To the Editor: In a recent issue of Critical Care Medicine, Ohannessian et al (1) estimated an average 5-day increase in ICU length of stay (LOS) due to hospital-acquired infections (HAIs) in a population of 33,449 patients from 12 French ICUs that were followed from 1995 to 2012. We join Lee and Bittner (2) in commending the authors for their study, which is not only unprecedented in size but also accommodates certain preventable yet common types of statistical bias. Their use of multistate models (MSMs) enables proper account of the timing of events, eliminating so-called time-dependent bias which arises when naively comparing the average LOS for infected and uninfected patients “as if patients’ future infection status were already known at ICU admission.” Such comparison, although common (3), erroneously apportions time-to-infection to prolonged LOS due to infection. In particular, it artificially prevents infected patients from dying or getting discharged before having acquired infection, leading to overestimation of excess LOS. MSMs not only enable adequate modeling of competing outcomes but also explicitly account for the timing of HAIs, thereby naturally accommodating the changing composition of infected and uninfected patients over time. Although the authors’ analysis made some adjustment for confounding by baseline covariates, patient conditions are highly dynamic, and admission characteristics, therefore, do not suffice for confounding adjustment (even more so when limited to just 18 strata). As of yet, suggested MSM approaches are not sufficiently tailored to accommodate confounding by time-dependent patient conditions, for which more advanced causal inference techniques are warranted (4). We are moreover, and especially, concerned about the reported dose-response relationship between the number of HAIs and excess LOS. We strongly believe that it exaggerates the impact of additional infections because of a remaining time-dependent bias. Whereas the authors correctly acknowledge that future infection status is unknown at admission Methylprednisolone: Likely an Effective Treatment in Acute Respiratory Distress Syndrome. Crit Care Med 2019; 47:e374–e375 2. Annane D, Pastores SM, Rochwerg B, et al: Guidelines for the diagnosis and management of critical illness-related corticosteroid insufficiency (CIRCI) in critically ill patients (Part I): Society of Critical Care Medicine (SCCM) and European Society of Intensive Care Medicine (ESICM) 2017. Crit Care Med 2017; 45:2078–2088 3. Annane D, Sébille V, Bellissant E; Ger-Inf-05 Study Group: Effect of low doses of corticosteroids in septic shock patients with or without early acute respiratory distress syndrome. Crit Care Med 2006; 34:22–30 4. Meduri GU, Golden E, Freire AX, et al: Methylprednisolone infusion in early severe ARDS: Results of a randomized controlled trial. Chest 2007; 131:954–963 5. Steinberg KP, Hudson LD, Goodman RB, et al; National Heart, Lung, and Blood Institute Acute Respiratory Distress Syndrome (ARDS) Clinical Trials Network: Efficacy and safety of corticosteroids for persistent acute respiratory distress syndrome. N Engl J Med 2006; 354:1671–1684 DOI: 10.1097/CCM.0000000000003648 (by modeling the first HAI as an intermediate state), they then conduct the dose-response analysis “as if the number of subsequent infections were known upon diagnosing the first infection” (by treating the number of HAIs as a time-fixed covariate), leading to overestimation of excess LOS due to multiple HAIs and underestimation of excess LOS due to a single HAI. By incorrectly apportioning patient-time contributions between first and subsequent infections to prolonged LOS due to additional infections in already infected patients, the authors inadvertently reintroduce the same type of bias they intended to eliminate. This could have been avoided by explicitly modeling each subsequent infection as an intermediate state. To illustrate, we replicated the authors’ flawed analysis in 2,728 ventilated patients (admitted to the Ghent University Hospital ICUs between 2012 and 2017) and compared results to those obtained by our suggested approach (unpublished findings). LOS increased by 4.5 or 11.3 days on average if uninfected patients had acquired either a single HAI or more than two HAIs, respectively, by the third ICU day (and no additional infections were acquired thereafter). In contrast, the authors’ approach would have led us to conclude that average LOS was prolonged by 3.6 or 33.3 days, respectively. Dr. Benoit’s institution received grants from Gilead, Astellas, Fisher-Paykel, Baxter, Alexion, and Fresenius Kabi outside the submitted work. The remaining authors have disclosed that they do not have any potential conflicts of interest.