Study design and settings
This is a retrospective observational study using data from the ICU of the Japanese Red Cross Maebashi Hospital from July 2013 to June 2017. The Japanese Red Cross Hospital ICU, which was 12-bed closed mixed ICU, had approximately 800 ICU admissions per year during the study period. The ethics committee of the hospital approved this study and confirmed that the need for informed consent was waived due to the retrospective nature of the study.
We included patients aged ≥ 18 years who met the diagnostic criteria for sepsis based on the Sepsis-3 criteria at the time of the ICU admission and stayed in the ICU for ≥ 48 h [24,25,26]. The eligibility of patients who were admitted to the ICU were retrospectively evaluated by two of the authors prior to applying the Sepsis-3 criteria . We excluded the following patients, because they were thought to have limited capacity to ambulate during their ICU stays: patients with acute cerebrovascular disease, progressive neuromuscular disease, post-cardiac arrest syndrome, unstable pelvic fracture, spinal injury with fracture of the spine, or multiple absent limbs. If a patient was readmitted to the ICU after discharge from the hospital during the study period, only data from the first admission was used for the analysis. All patients received the standard treatment based on surviving sepsis campaign guidelines 2012  and 2016 .
The Maebashi early mobilization protocol
In this study, mobilization was defined as rehabilitation at the level of sitting on the edge of the bed or more (e.g., standing beside or walking around the bed). At Maebashi Red Cross Hospital, there was no standardized protocol for the introduction of mobilization for patients admitted to the ICU, but in June 2015, the Maebashi early mobilization protocol was created. Details of the Maebashi early mobilization protocol are provided in Additional file 1. Although the Maebashi early mobilization protocol changed the timing of mobilization introduction, it did not change the mobilization level that was provided to the patients and the 20-min duration of mobilization per session. The discontinuation criterion at each rehabilitation session was described in Additional file 1, in line with recent expert consensus .
The data were retrospectively collected from electronic-based medical records . We collected the following patient demographics and characteristics: age, sex, body mass index (BMI), Charlson Comorbidity Index (CCI), APACHE II and SOFA score at ICU admission, the main source of infection, the route to the ICU (e.g., emergency room, general ward), the ambulatory dependence before hospital admission, the diagnosis of septic shock at ICU admission, and the receipt of the Maebashi early mobilization protocol. We also collected data on when patients first received rehabilitation interventions and when patients first achieved mobilization during their ICU stays. In addition, we collected the data on the treatment patients received during their ICU stays: the use of the medical devices (invasive mechanical ventilator, extracorporeal membrane oxygenation [ECMO], and renal replacement therapy), corticosteroids, neuromuscular blockade, continuous analgesia with fentanyl, continuous sedation with benzodiazepines, propofol, or dexmedetomidine, and continuous vasopressor infusion (norepinephrine, dopamine, dobutamine, epinephrine, or vasopressin). For continuous analgesia, sedation, and vasopressor use, the details of the name of the drug used, its duration, and the average doses were also collected.
The primary outcomes were in-hospital mortality and ambulatory dependence at hospital discharge. The secondary outcomes were the lengths of the ICU and hospital stays and the total hospital costs. The total costs were calculated based on the Diagnosis Procedure Combination/Per-Diem Payment System  and converted from Japanese yen to US dollars at an exchange rate of 114 yen/dollar.
First, we defined early mobilization (EM) as achieving mobilization within the first 3 days of ICU stay. Those who did not achieve mobilization during their ICU stay or achieved mobilization after the first 3 days were classified into the non-EM group. We compared the patient characteristics, treatments, and outcomes between the two groups by using the Mann–Whitney U-test and the Fisher’s exact test.
Second, we developed a multivariable logistic regression model to estimate a propensity score for each patient’s likelihood of achieving early mobilization. The covariates included to generate the propensity score were as follows: age, sex, BMI, CCI, APACHE II and total SOFA score at the ICU admission, the route to the ICU, ambulatory dependence before the hospital admission, the diagnosis of septic shock at ICU admission, the receipt of the Maebashi early mobilization protocol, and the treatments which patients received during their ICU stays [invasive mechanical ventilation, ECMO, renal replacement therapy, corticosteroid, neuromuscular blockade, analgesia with fentanyl, sedation with midazolam and propofol, and catecholamine use (noradrenaline, dopamine, or dobutamine)]. We then applied stabilized inverse probability weights (sIPWs) [32, 33] to calculate the adjusted odds ratios (OR) and 95% confidence intervals (CIs) of patients in the EM group relative to the non-EM group for the primary outcomes, and the adjusted means and 95% CIs for both groups for the secondary outcomes.
Next, we further analyzed data using alternative definitions of EM by changing the cutoff in days to mobilization by 1-day increments from 2 to 7 days. For each definition of EM, we implemented applied sIPWs as aforementioned to examine the changes in outcomes.
In addition, we performed two subgroup analyses: (i) excluding patients who did not achieve mobilization during their ICU stay from the non-EM group and (ii) excluding patients before June 2015, when the Maebashi early mobilization protocol was introduced. In each analysis, similar to the main analysis, we changed the cutoff in days to mobilization by 1-day increments from 2 to 7 days and applied sIPWs as aforementioned.
All analyses were conducted using Python version 3.8.12. A P-value of < 0.05 was considered statistically significant.