We conducted a population-level, observational, retrospective study using prospectively gathered data from patients admitted to 168 ICUs in ANZ between January 2017 and September 2020 contained in the ANZ Intensive Care Society Adult Patient Database (ANZICS APD). Patients aged > 16 years at time of ICU admission were included. We excluded patients admitted solely for organ donation or palliative care, those transferred between ICUs (with uncertain duration of total ICU exposure), or those re-admitted to ICU during the index hospitalisation.
Clinical data collectors measured frailty on ICU admission using the Canadian Study of Health and Aging Clinical Frailty Scale (CFS) [
10]. The CFS is a judgement-based categorical scale based on patients’ baseline fitness, which correlates well with the original 70-item frailty index used in this original study. It is also valid and reliable when applied to a range of acutely ill populations, and is the dominant frailty scale used in ICUs worldwide [
8,
11‐
13]. The CFS is modified in the ANZICS-APD to eight categories: CFS = 1 (very fit), CFS = 2 (well), CFS = 3 (managing well), CFS = 4 (vulnerable), CFS = 5 (mildly frail), CFS = 6 (moderately frail), CFS = 7 (severely frail), or CFS = 8 (very severely frail). Terminally ill patients, usually scored 9 on the CFS, are instead scored in the APD on their level of frailty. The CFS was analysed in four categories (CFS 1–2, 3–4, 5–6, 7–8), and also dichotomised (frail: CFS = 5–8, non-frail: CFS = 1–4). The study was approved by the Alfred Hospital Human Research Ethics Committee, individual patient consent was not required (HREC-ref. 584/18). During the study period of interest (January 2017–September 2020), 196 ICUs from Australia and New Zealand contributed data to the APD representing more than 90% of all ICU admissions during this period. A total of 86% (168/196) of these sites contributed patient information pertaining to frailty.
Statistical analysis
The primary outcome was in-hospital mortality during index hospitalization. Exposure variables were frailty, calculated using the CFS, and risk of death predictions calculated separately using antecedent characteristics or acute illness components as previously described [
1]. The antecedent characteristics included features related to the patient (age, smoking status, comorbidities, treatment limitations), the admitting ICU (location, size, type) and temporal trends relating to the timing of admission (hour, day, month, and year).
The acute illness prediction model included features collected in the 24 h following ICU admission and include Acute Physiology and Chronic Health Evaluation (APACHE) admission diagnosis; APACHE III acute physiology scores; ICU admission source (emergency department, operating theatre, ward); ICU care type (ICU vs high-dependency unit); pre-ICU length of hospital stay; mechanical ventilation; medical emergency team call, respiratory arrest, or cardiac arrest in the previous 24 h. Models were calculated using parameter estimates outlined above derived from 514,117 patients included in the 2000–2014 cohort in which PerCI was first described [
1]. Primary multivariable analyses for in-hospital mortality were conducted using logistic regression, individually analyzing patients still in ICU with separate regression models conducted each day between day 1 (the day of ICU admission) and day 21. To enable prediction models to be constructed using all available data, where data was missing for categorical variables, an additional category was created and fitted to account for missingness. Where continuous physiological variables were missing, single imputation with normal value substitution was conducted in accordance with standard clinical risk modelling [
14]. No imputation was performed for the primary exposure variable (frailty).
The contribution of acute and antecedent characteristics to in-hospital mortality risk prediction were examined via differences in the area under the receiver operating characteristics (AUROC) curve for each regression, with statistical comparison performed using chi-square tests. To establish the increased risk of death associated with frailty, patients from the highest frailty category (CFS 7–8) were compared against patients from the lowest frailty category (CFS 1–2), adjusting for both acute and antecedent characteristics, with results presented as odds ratios (95% CI). Two additional sets of models were further constructed to determine the independent discriminatory capacity of frailty to predict in-hospital mortality. These models include the following: (a) incorporating the antecedent risk of death prediction, the acute illness risk of death prediction, and an interaction term between the two, and (b) the same model with the addition of frailty status, and interaction terms between frailty and the existing model variables as defined in (a). To further model the effect of frailty-data missingness, a sensitivity analysis was conducted on a subset of ICUs that had high completion (> 80%) of frailty data for the entire study period.
Multivariable analysis for the prediction of PerCI (a priori defined as an intensive care stay in excess of 10 days) was performed using logistic regression adjusting for characteristics that have been previously identified as being associated with PerCI [
1]. Variables included in the model pertain to the ICU (location, type and size), time (hour, day and year), patient (frailty, age, gender and comorbidities), ICU admission (care type, source, treatment limitations and Pre-ICU length of stay) and patient severity (admission diagnosis and physiological derangement). Results are presented as odds ratios (95% CI). Raw comparison of proportions were compared using chi-square tests for equal proportion.
All analysis was performed using SAS version 9.4 (SAS Institute Inc., Cary, NC, USA) and to increase the robustness of our analysis a two sided P value of 0.001 was used to indicate statistical significance.