Method of Measuring Comorbidity to Predict Outcome After Intensive Care

Overview

In this study the investigators will validate the impact of comorbidity on readmission to intensive care unit (ICU) and mortality after ICU and which method of measuring comorbidity that is most predictive. The study population included all critical care patients' registries in Swedish intensive care registry (SIR) during the years 2005 to 2012 with valid personal identity number. Data from Statistics Sweden och National Board of Health and Welfare were linked to data from SIR and de-identified. Hospital discharge diagnoses from five year preceding the index date for the ICU admission were extracted. A composite outcome of death and readmission will be analyzed. Analyzes with cox proportional-hazards regression, time to event, on the training data set year 2005-2010 The study population will be split in a training data set (2005-10) and a test data set (2011-12) for validating our prognostic model. The predictive ability in the test data set were evaluated based on discrimination, AUC (C index), Calibration and Brier score.

In this study the investigators will validate the impact of comorbidity on readmission to intensive care unit (ICU) and mortality after ICU and which method of measuring comorbidity and in which time-point it is most predictive. The study population includes all critical care patients' registries in Swedish intensive care registry (SIR) during the years 2005 to 2012 with valid personal identity number. SIR delivers the population to Statistics Sweden directly or via the client for further delivery to Statistics Sweden. For all individuals in the population, data are collected from registers at Statistics Sweden. Statistics Sweden supplies social security numbers and serial numbers to the National Board of Health for further data collection there. The National Board of Health and Welfare delivers the data sample to the client who sends it to Statistics Sweden for collaboration and anonymous. All data material is stored unidentified in the MONA database where only persons connected to the project have access to the material. The data set containing 293 342 observations and 223 495 unique individuals. Observations which is totally covered in time by another observation are excluded. Two consecutive observations with less than 24 hours between them consider as the same visit. The final set of dates consists of 273 741 observations (223 495 individuals). Patients with recurrent ICU stays during the study period were considered as recurrent events that are not independent of each other. The interval between ICU discharge and readmission was used both as an outcome variable and to characterize the patient at the time to admission. In the analyzes the dependency between multiple admissions for the same individual was handled using a robust sandwich estimator. Every ICU stay was included in the study but handled as a time-updated exposure. A composite outcome of death and readmission will be analyzed. Death and readmission will also be analyzed separately. Follow-up starts at admission. A binary status variable (no/yes) is created reflecting if the outcome has happened or not together with a corresponding time variable. For each admission the follow-up ends with readmission, death or end of study (2016-12-31) whichever comes first. Hospital discharge diagnoses from five years preceding the index date for the ICU admission were extracted from the National Board of Health and Welfare and linked to the SIR data using exact person-based linkage. The Charlson comorbidity index was calculated based on this information. A different categorization of comorbidity was also performed as modified based from the categorization proposed by Elixhauser. For each of 36 defined comorbidity categories the number of admissions with a primary diagnosis, the number of admissions with a secondary diagnosis, the total length of stay with a primary diagnosis, and the interval from the last admission with the comorbidity condition as a primary diagnosis, were calculated. The underlying condition causing ICU admission was categorized according to diagnosis and admitting department. Analyzes with cox proportional-hazards regression, time to event, on the training data set year 2005-2010 Charlson comorbidity index (CCI) categorical 0, 1, 2, 3-5, 6-9, 10-17 Elixhauser 36 categories 1. Number of primary diagnosis (count) 2. Number of secondary diagnosis (count) 3. Total care time primary diagnosis (count) 4. Time interval from latest primary diagnosis (missing, count) Analyzes A Age + sex B A + CCI C A + a D A + a + b E A + a + b + c F A + a + b + c + d There was by definition no missing data in the comorbidity variables. Missing information concerning age and sex were minimal and did not require imputation. The proportional hazards assumption was checked using the Kaplan-Meier Curves. The study population will be split in a training data set (2005-10) and a test data set (2011-12) for validating our prognostic model. The predictive ability in the test data set were evaluated based on discrimination, area under curve (AUC) (C index), Calibration and Brier score.