Study population
At the population level of one French region, all adult patients admitted to an ICU from January 1st, 2010 to December 31st, 2012, for septic shock and/or ARDS, with 5 days or more invasive mechanical ventilation, were extracted from the regional medico-administrative database (“Programme de Médicalisation des Systèmes d’Information”, PMSI) using a computerized algorithm. This database relies on the mandatory notification of each hospital stay, through a coded summary, for all public and private French hospitals. Every hospital stay in the database is linked to patient data using an encrypted anonymized number, allowing to carry out epidemiological analyses among a comprehensive historical cohort. The study was performed in one representative French region (Centre Val de Loire, 2.5 million inhabitants), including one university hospital, one tertiary hospital and 37 general and private hospitals. Eight hospitals of the region have at least one intensive care unit (Fig. 1).
Algorithm selection criteria were (see Additional file 1: Table S1, Additional file 2: Table S2):
Presence of “ARDS” and/or “Septic shock” codes from the International Classification of Disease, Tenth Revision (ICD-10), as primary or secondary diagnosis.
Invasive mechanical ventilation reported for 5 days or more, using the corresponding codes from the French Common Classification of Medical Acts.
Patient under 18 years and pediatric ICUs were not included.
Performance of the selection algorithm was validated by a blinded review of medical charts of 180 patients randomly selected, comprising 70 cases selected by the inclusion algorithm and 110 controls, from two different hospitals and admitted in ICU during the study period and who did not met criteria of the selection algorithm. Investigators then blindly reviewed medical chart for case validation. ARDS was defined according to the Berlin Definition [20], and septic shock was defined as the presence of a suspected or proven infection associated with hypotension requiring vasopressive therapy after adequate fluid resuscitation (as mentioned in the medical chart, and/or at least 30 mL/kg). Invasive mechanical ventilation was defined by mechanical ventilation through a tracheal tube or tracheostomy. Algorithm performance appeared excellent with a positive predictive value of 96%, and a negative predictive value of 92%.
The purpose of the study was to explore post-ICU burden in terms of healthcare use; thus, among ICU patients identified by the algorithm, only patients discharged alive from the hospital were included in the study cohort.
Data collection
Index hospitalization was defined as the whole acute care hospital stay during which the index ICU stay occurred. If a patient had two or more hospitalizations satisfying inclusion criteria, the first hospitalization was considered as the index hospitalization. Age, sex, simplified acute physiology score (SAPS) 2, duration of mechanical ventilation, length of stay in the ICU and overall hospital length of stay were collected for each index hospitalization. Patients were tracked during a 2-year period before and a 2-year period after the index ICU hospitalization (which are, respectively, termed “pre-ICU period” and “post-ICU period” thereafter).
Comorbidities were retrieved for each patient. For this purpose, we extracted ICD-10 coding related to chronic comorbidities and grouped them into categories according to the Charlson Comorbidity Index [21] (Additional file 3: Table S3). Comorbidity-related ICD-10 codes were then tracked during the pre-ICU and the post-ICU periods to build pre-ICU and post-ICU comorbidity reports. Mortality at hospital during the post-ICU follow-up was also recorded.
Healthcare use was similarly analyzed during pre-ICU and post-ICU period using the exact same methodology for data extraction. Specifically, we extracted the number of days in hospitalization and the number of ambulatory care/consultations (considered as 1-day healthcare use). We also extracted the type of healthcare use: acute care settings (either medical or surgical), rehabilitation centers, psychiatry, or “hospitalization at home” (a specific setting with high intensity care and nursing organized at the patient’s home).
Analyses
To assess consequences of active medical conditions in terms of healthcare use, we specifically focused on the recourse to acute care facilities, either for complete hospitalizations (at least one overnight stay) or for ambulatory hospitalization (full day spent in hospital for multiple consultations, interventions, and treatment) that we designated as “days of healthcare use”. Thus, hospitalization in rehabilitation centers and simple ambulatory consultations were not included in this definition.
Temporal trends of healthcare use analyses were performed by quarterly assessing frequency of days of healthcare use (% of days), calculated as the quarterly report of the numbers of days of healthcare use divided by the number of patient-days during the quarter. For instance, a patient who would have spent 5 days at a hospital, and then 4 ambulatory hospitalizations during a quarter would be considered as having 9 days/90 days = 10% of days spent for healthcare use. This procedure enables to take into account patients lost for follow-up and deaths in the post-ICU period.
The main hypothesis underlying the present work was that the population of patients admitted in ICU for ARDS/septic shock would have heterogeneous healthcare trajectories prior to ICU admission, and thus, further analysis of post-ICU healthcare use should integrate this parameter.
To do so, we had to group patients according to (1) their level of pre-ICU healthcare use, and (2) the temporal dynamic of this pre-ICU healthcare use. Practically, this could be done by clustering patients according to quarterly evaluation of healthcare use in pre-ICU period. We used the K-Means clustering technic, allowing an unsupervised and unbiased approach to aggregate patients together based on their similarity in healthcare use for each quarter in pre-ICU period. For this purpose, and to capture all aspects of healthcare use, we used a broader definition of healthcare use for the clustering, encompassing recourse of hospitalizations, ambulatory consultations, in acute care settings and rehabilitation centers. We tested the algorithm with a number of clusters ranging from 2 to 10 (i.e., K from 2 to 10) and analyzed for each test the percentage of inter-group variance explained by constitutions of the groups. When increasing the number of clusters (i.e., increasing the value of k), the inter-group variance explained by those clusters likely increases. Thus, the objective of this method is to reach a reasonably high inter-group variance explained by constitution of the clusters (i.e., close to 70%), with a clinical relevant number of clusters. At the end, in our dataset, a K = 5 led to a good variance (67%) along with a relevant number of clusters to analyze, and with no further significant increase in inter-group variance for K > 5. Once clusters of patients were built, we could next plot the quarterly healthcare use—in both pre- and post-ICU periods—for each cluster to build pre- and post-ICU healthcare trajectories in the different clusters.
For pre-post ICU periods comparisons, Wilcoxon and McNemar tests for paired data were used, as appropriate.
For comparisons between groups, Chi-square and ANOVA tests were used, as appropriate.
Kaplan–Meier survival curve analysis was used to explore mortality during the post-ICU period.
Statistical analyzes were performed with SAS version 9.4 (SAS Institute Inc., Cary, NC) and R 3.2.2 (https://www.R-project.org). A p < 0.05 was as considered statistically significant.