Design, Setting, Participants, Data Sources and statistics
For this observational study, we have used a high-resolution cohort described in details elsewhere [2]. Briefly, we collected data from all adult (> 18 years old) patients admitted to our tertiary ICU between January 1st 2010 and June 15th 2020. We excluded patients with documented or expressed wishes of non-participation to clinical research, those with end-stage renal disease (ESRD), whose stay lasted for less than 6 hours or with missing outcome or AKI defining data. For each patient, we considered only the first eligible ICU admission. Data were extracted from electronic medical records [Metavision®(IMD Soft, Tel Aviv, Israel) and Soarian®(Cerner, North Kansas City, USA)]. In particular, we collected baseline characteristics, comorbidities, pre-admission weight, illness severity scores and hourly UO measurements. The primary outcome was 90-day mortality. Patients' vital status was assessed by cross-referencing our dataset with the Swiss national death registry.
Continuous data are reported as mean (standard deviation, SD) or median (interquartile range, IQR) according to underlying data distribution. Categorical variables are expressed as number (percentage). All statistical analyses and modelling were carried out in R version 4.1.2 [13]. The level of statistical significance was set at 5%.
Average standardized urinary output calculation
We standardized hourly UO (ml/h) (\({v}_{is}\)) by pre-admission body weight (\({w}_{i}\)) when available (missing values were imputed using multiple imputations see below). We then used a sliding window to calculate the average standardized urinary output (\({\overline{v}}_{it}\)) of a patient i over \(d\) hours preceding time t, such that
$${\overline{v}}_{it}\left(d\right)=\frac{1}{d}{\sum }_{s=t-d+1}^{t}\frac{{v}_{is}}{{w}_{i}}$$
For convenience, we restricted the analysis to sliding windows of width \(d=\left\{\text{3,6},\text{12,24}\right\}\) hours. Note that \({\overline{v}}_{it}\left(d\right)\) cannot be calculated when \(t<d\). We then calculated the minimum value among all moving averages of width d that can be computed over the whole ICU stay of each patient, that is
$${u}_{i}\left(d\right)=\underset{t\ge d}{min}\left({\overline{v}}_{it}\left(d\right)\right)$$
Hence, \({u}_{i}\left(d\right)\) corresponds to the minimal average UO that patient i experienced over a period of d hours during his ICU stay.
Of note, we considered pre-admission body weight when available.
Modelling 90-day mortality
Logistic regression was used to predict 90-day mortality as a function of the minimum average urine output of patients, separately for medical, scheduled surgical and unscheduled surgical admissions. Within each admission type, control variables included patient’s age at ICU admission, SAPS II score (corrected to not account for daily UO) and Charlson comorbidities index. All predictors were continuous and flexibly modelled using penalized thin plate regression splines within the framework of generalized additive models [14, 15]. Alternative candidate models also included smooth terms for the interaction between minimal average UO and other continuous predictors. A model which did not include minimum average UO as predictor (“base model”) was also fitted for comparative purposes. The model featuring the highest overall calibration performance (i.e. lowest mean squared error of prediction) across all time windows was selected using 10-fold cross-validation. For each value of d, predicted 90-day mortality was plotted as a function of the minimal average UO for fixed covariate patterns (e.g. corresponding to median values of control variables). This allowed visual identification of suitable oliguria thresholds i.e. UO thresholds below which mortality increases substantially, and compare these with thresholds used in current practice.
Data were randomly split into a training (80%) and a validation set (20%). Training data were used to develop and fit prognostic models (including selection of the best model using 10-fold cross-validation). Validation data were exclusively used to evaluate the final models’ discrimination and calibration properties. Discrimination was assessed using the area under the receiver operator characteristic curve (AU-ROC). Calibration was assessed using Hosmer-Lemeshow test [16] and calibration belt [17]. Discrimination and calibration performances were assessed on all validation data as well as on subsets of patients whose minimum average urine output fell below proposed thresholds for oliguria.
Handling missing covariate data
Missing covariate values were imputed using multiple imputations [18] with a total of 50 complete datasets being reconstructed. Multiple imputations were carried out separately within training and validation sets. Additional variables such as gender, body weight, baseline creatinine, need for noradrenaline were used in the imputation process. Note that when body weight \({w}_{i}\)was missing, imputed weights were used to calculate imputed values for \({\overline{v}}_{it}\left(d\right)\) and thus \({u}_{i}\left(d\right)\) as well. Rubin rules[18] were used to pool predicted mortalities, AU-ROC estimates and/or calibration results obtained within each complete dataset.
Urinary output is collected manually on an hourly basis by ICU nurses. The management of missing UO values is described in details in the supplementary material.
Ethics
This study was approved by the Ethics Committee Vaud (CER-VD 2017-00008, Lausanne, Switzerland). In accordance with the Swiss Federal Act on Research involving Human Beings (article 34)[19], retrospective utilization of non-genetic health-related personal data was permitted, provided that the patient (or its legal representative) had not expressed wishes of non-participating to clinical research. This study followed “The Strengthening the Reporting of Observational Studies in Epidemiology” (STROBE) guidelines for reporting observational studies.