Evaluation of a minimal sedation protocol using ICU sedative consumption as a monitoring tool: a quality improvement multicenter project

Introduction Oversedation frequently occurs in ICUs. We aimed to evaluate a minimal sedation policy, using sedative consumption as a monitoring tool, in a network of ICUs targeting decrement of oversedation and mechanical ventilation (MV) duration. Methods A prospective quality improvement project was conducted in ten ICUs within a network of nonteaching hospitals in Brazil during a 2-year period (2010 to 2012). In the first 12 months (the preintervention period), we conducted an audit to identify sedation practice and barriers to current guideline-based practice regarding sedation. In the postintervention period, we implemented a multifaceted program, including multidisciplinary daily rounds, and monthly audits focusing on sedative consumption, feedback and benchmarking purposes. To analyze the effect of the campaign, we fit an interrupted time series (ITS). To account for variability among the network ICUs, we fit a hierarchical model. Results During the study period, 21% of patients received MV (4,851/22,963). In the postintervention period, the length of MV was lower (3.91 ± 6.2 days versus 3.15 ± 4.6 days; mean difference, −0.76 (95% CI, −1.10; −0.43), P <0.001) and 28 ventilator-free days were higher (16.07 ± 12.2 days versus 18.33 ± 11.6 days; mean difference, 2.30 (95% CI, 1.57; 3.00), P <0.001) than in the preintervention period. Midazolam consumption (in milligrams per day of MV) decreased from 329 ± 70 mg/day to 163 ± 115 mg/day (mean difference, −167 (95% CI, −246; −87), P <0.001). In contrast, consumption of propofol (P = 0.007), dexmedetomidine (P = 0.017) and haloperidol (P = 0.002) increased in the postintervention period, without changes in the consumption of fentanyl. Through ITS, age (P = 0.574) and Simplified Acute Physiology Score III (P = 0.176) remained stable. The length of MV showed a secular effect (secular trend β1 = −0.055, P = 0.012) and a strong decrease immediately after the intervention (intervention β2 = −0.976, P <0.001). The impact was maintained over the course of one year, despite the waning trend for the intervention’s effect (postintervention trend β3 = 0.039, P = 0.095). Conclusions By using a light sedation policy in a group of nonteaching hospitals, we reproduced the benefits that have previously been demonstrated in controlled settings. Furthermore, systematic monitoring of sedative consumption should be a feasible instrument for supporting the implementation of a protocol on a large scale. Electronic supplementary material The online version of this article (doi:10.1186/s13054-014-0580-3) contains supplementary material, which is available to authorized users.


Data analysis
To analyze the effect of the campaign, we planned a quasi-experimental design using interrupted time series (ITS) in order to control for secular trends [1][2][3][4]. In an ITS design, data are collected at multiple instances over time before and after an intervention (interruption) and is the strongest, quasi-experimental design to evaluate effects of time-delimited interventions. An advantage of ITS design is that it allows investigation of potential biases further them the secular trend, as duration of the intervention (i.e. the intervention might have an effect for the first two months only and not a sustained one).

Interrupted time series analysis and ARIMA models
To investigate such biases, we run autoregressive integrated moving average models (ARIMA). Two parameters define each segment (period before and after the intervention) of a time series: level and trend. The level is the value of the series at the beginning of a given time interval. The trend is the rate of change of a measure (slope) during a segment. To examine the results, we might analyze if there are changes in level and trend that follow an intervention. In general, a change in level constitutes an abrupt intervention effect and a change in trend represents a gradual change in the value of the outcome. This model involves a multilinear regression and we can specify the following linear regression model to estimate the level and trend for our primary endpointlength of mechanical ventilation: Below is the hypothetical data to exemplify how to organize the data to conduct an interrupted time series analysis. In this short example, the intervention occurred after 5 months of the observed period. Time can be seconds, hours, days, months, and years; however it must be at regular intervals for this classical analysis. The assumptions required for ARIMA models, the stationary process which implies that the mean and variance do not change over time, were checked using Phillips-Perron, the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) and Augmented Dickey-Fuller tests. The autocorrelation was checked by visual inspection of autocorrelograms and partial autocorrelograms of the series and its residuals. The White neural network test was used to test for neglected nonlinearity. We also checked for seasonal or cyclical effects by decomposing our series and any of these patterns were detected. The Ljung-Box Q test was run to evaluate a lack of fit of the final ARIMA model.

Hierarchical time series
To fit the hierarchical time series [5] through a bottom-up method, first we get independent time series at the bottom level of the hierarchy (each ICU, Level 1) and then aggregating the independent time series upwards to produce a revised time series for the whole hierarchy. This method accounts for noise and variability between time series and provides additional information in comparison of crude average.
In our study, the Level 0 is the most aggregated level and represents the network level. As depicted in the scheme above, the Level 0 is composed by 10 series at Level 1. The Y0,t is the tth observation of the Level-0 series for t=1,2,3,4,…, 24 (in our study). For Level-1, we denote Y1,t the tth observation of the series at level 1 (Unit level). The total amount of times series is n = 1 + 10 = 11. To represent the aggregation of the observations, the hts package used a matrix notation, constructing a n x nk matrix, where nk is the number of time series at the lowest level of the hierarchy.
The hierarchy can be represented by: Unit 08 Level1 fit an interrupted time series following the previous statements. All the background explicit above and analyses were retrieved through the hts package v 4.4 from professor Rob J Hyndman and coworkers [5].

Generalized Linear Mixed Model
To deal with variation in case-mix over time at two levels (patient and unit), we conducted a sensitivity analysis for our main outcome fitting a Generalized Linear Mixed Model through the segmented regression concept, achieving the better fit for correlated responses, in particular for the analysis of our longitudinal and clustered data. Length of MV for each patient was the dependent variable. For the first level (patient level), we adjusted for SAPS-3, Charlson score and vasoactive drugs.
For the second level (unit level), we fit random intercepts for each one of 10 units and nested into units variable random components for types and reasons for admission: sepsis syndrome, cardiac surgery and respiratory reason. The model was built with a Poisson distribution with Log as a link function and the covariance structure for the random effects was the AR1 (Firstorder autoregressive) [34,35]. In this model, patient-level was constituted by fixed effects. For the unit-level, the factors were fit as random effects. Therefore, we can model the individual differences between units by assuming different random intercepts for each unit.
This analysis was conducted through the lme4 package and the glmer function. The R code for random effects can be defined as (1|units), when the 1 requires an intercept and the units variable insert a random effect into the model for each unit.
We also nested the above variables per units, which the code is (1|units/sepsis), where factor sepsis is nested in units, for example. To model the within-group correlation not captured by the random effects, we fit the model with a correlation structure as AR1, by the "cor=corAR1" term.
In order to assess the contribution of the random effects on the model we ran two evaluations. One of them was plotting the random intercepts with the error around each estimate. These parameters were retrieved through the ranef function on the lme4 package [8]. Our efigure7 is an example. Another approach was evaluating the variance parameter for random effects, which can represent the percentage of the model explained by the random components [8,9]. The "interclass correlation coefficient", used for linear mixed models, was adapted and calculated for generalized linear mixed model as described by Goldstein [10]. Therefore, we were able to retrieve the amount of the model attributed to the random effects. To test whether the random components resulted on better overall model fit in comparison to the model without the random component, we ran a Likelihood ratio test [8,9].