In brief, CHEST was a multicenter, prospective, randomized-controlled, clinical trial (RCT) that compared the use of HES and saline for intravenous fluid resuscitation in patients in 32 ICUs in Australia and New Zealand (Clinicaltrials.gov identifier#: NCT00935168). Detailed descriptions of the study protocol, statistical analysis plan and results have been published previously (16, 17). The CHEST study was approved by the human research Ethics Committee of Northern Sydney Central Coast Health (AU RED Ref: HREC/09/HARBR/14), and by each participating institution. The present study is a post-hoc analysis of the CHEST database with the objective of developing a model for the prediction of severe AKI in ICU patients receiving fluid resuscitation.

The primary outcome of the model is severe AKI, defined as the first event in a composite outcome incorporating doubling of serum creatinine (from the pre-randomization value) and/or treatment with RRT within 28 days of randomization. It is important to note that this outcome differs from the reported AKI outcome in the primary CHEST article, which used the RIFLE criteria (17) defined by changes in either urine output or serum creatinine from randomization.

In order to minimize selection bias, our analysis included all patients from the original study population for whom consent to use of their data was obtained and were not lost to follow-up. Demographic and clinical variables were collected as described previously (16), and variables collected immediately before randomization were deemed as baseline variables. We examined 17 candidate baseline variables that were prospectively considered by the authors to be potential indicators of AKI risk: age, sex, weight, heart rate (HR), central venous pressure (CVP), mean arterial pressure (MAP), urine output during 6 h prior to randomization, serum creatinine (serum Cr), estimated glomerular filtration rate (eGFR), serum lactate concentration, Acute Physiology and Chronic Health Evaluation (APACHE) II score (18), Sequential Organ Failure Assessment (SOFA) score for the cardiovascular system (19), the presence of sepsis (20), the presence of trauma, source of admission, treatment with mechanical ventilation, and the nature of ICU admission (surgical or nonsurgical). As the allocated study treatment (HES or saline) was assigned at randomization, this variable was not considered as a baseline variable and excluded from multivariate regression models.

eGFR was calculated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation (21). The source of admission was defined as the place from where the patients had been transferred to ICU and was classified as: hospital floor, emergency department, operating room following elective surgery, operating room following emergency surgery, and other hospitals. Mechanical ventilation included either invasive ventilation via an endotracheal tube or non-invasive respiratory support via a mask or other non-invasive interface.

As the analysis focused on the baseline clinical variables, all patients in the baseline analysis of the original study publication with complete demographic data and information about the study outcome were included. Continuous variables were presented as median (interquartile range) or mean ± standard deviation (SD), categorical variables were presented as number (percentage). APACHE II score was considered as a continuous variable. Univariable regression was used to examine the relationship between candidate predictors and the study outcome. Multivariable logistic regression model was used to identify independent risk factors for the study outcome.

For modeling, we followed a three-step procedure. First, we developed a primary model by multivariable logistic regression using a backward elimination approach (threshold p < 0.05) to select variables (22) after excluding variables that would potentially have high risk of in the development population. Possible first-order interactions were also explored unavailability in real-world clinical practice based on their proportions of missing data and kept in the model if statistically significant (p < 0.05). Second, we developed a secondary model by removing interaction terms in the primary model and compared the performance of these two models using c-statistics (equal to the area under the receiver-operating characteristic curve). If no difference between these primary and secondary models was found, the secondary model was preferred due to its greater simplicity and avoidance of over-fitting, consistent with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statements (23).Third, we simplified the model from step two by removing variables that were not statistically significant and tested the performance of the final parsimonious model. The three models were compared, and the final model was chosen as the best balance between calibration, discrimination and clinical practicability. To test the performance of the model, discrimination was assessed by c-statistics (24), and calibration was assessed by Hosmer-Lemeshow test (25) and calibration plot.

In addition we performed two sensitivity analyses. The first tested whether forcing the randomization treatment (HES or normal saline) into the multivariable regression model, altered the model's predictive performance, and the second examined whether the timing of the primary outcome, at 7 days rather than 28 days from randomization altered model performance to be consistent with current diagnostic criteria of AKI.

For validation, we randomly sampled from the study dataset using the bootstrapping method (26) (n = 10,000 replications) to evaluate the over-optimism inherent in the final model. Confidence intervals of the over-fitting in c-statistics for the bootstrap corrections were computed based on the assumption of normal distributions. The bootstrap-adjusted performance was calculated and the final risk prediction model was translated into a predictive equation. Positive predictive value (PPV) and negative predictive value (NPV) were calculated, and a decision curve analysis was performed to evaluate the net benefit of using the model in predicting the primary outcome across a range of threshold probabilities.

The reporting of this prognostic model study followed the TRIPOD statement (23). The risk of bias of the final model was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST) (27). In addition, we used the final model and REDCap electronic data capture tools (28) securely hosted at Sichuan Provincial People's Hospital to develop a web-based risk assessment tool accessible to readers.

All analyses were performed using SAS/STAT software v.9.1 (SAS Institute Inc., Cary, NC, USA) with statistical significance set at P < 0.05.

Baseline variables and the study outcome in the study population are listed in Table 1. Of the 6,742 patients reported in the original publication, 15 patients were excluded because of missing RRT follow up data, leaving 6,727 patients for analysis. The majority of the 6,727 patients entered ICU from the operating room or emergency department. Data for urine output during 6 h prior to randomization, baseline CVP and baseline serum lactate were available for only 2,802, 2,300 and 5,555 patients respectively, so these three variables were excluded from modeling (see Table 1). The proportion of individuals with at least one variable with missing value was 2.79% (188/6727).

Within 28 days after randomization, 514 (7.6%) participants experienced a doubling of serum Cr, 427 (6.4%) were treated with RRT and 196 (2.9%) experienced both events. In total 745 (11.1%) participants developed the study outcome (Figure 1).

The univariate relationships of all candidate model variables with the study outcome are presented in Table 2. The majority of these variables were associated with the study outcome, but sex, MAP, admission following emergency surgery (compared to admission from emergency department) and mechanical ventilation were not.

Evaluated variables in each model were examined and compared (Supplementary Table S1). In addition to stand-alone significant variables including age, baseline eGFR, heart rate, APACHE II score, sepsis, mechanical ventilation at admission, and admission source, the primary model (Model A) also included eight interactive terms as significant factors. However, the performance of the secondary model (Model B) after removing interactive terms did not materially differ from the primary model (Figure 2). After removing these interactive terms, seven variables became non-significant, resulting in the final model (Model C) including six significant predictors (Table 3). The comparison of all three models indicated no meaningful difference in C-statistics for the primary, secondary and final models (Figure 2).

The sensitivity analysis of including randomization treatment in the final model showed no change in the effect of the six predictors or the C-statistic (Table 4). The sensitivity analysis of the ability of the final model to predict doubling of serum creatinine and RRT within 7 days after randomization again showed no change in the effect of the model variables indicated all predictors remained significant except for mechanical ventilation at admission (Supplementary Table S2).

The observed and predicted risk of the composite primary outcome based on the final model were similar, with a test of goodness of fit indicating good calibration (modified Hosmer-Lemeshow test, χ² = 14.4, p = 0.07) (Figure 3). Internal validation using the bootstrap method revealed the degree of over-optimism on c-statistics of the final prediction model was 0.0055 (95% CI, −0.0129 to 0.0240), resulting in an equivalent c-statistic after bootstrap validation of 0.711. The PPV and NPV of the final model to predict severe AKI in ICU patients receiving fluid resuscitation within 28 days from admission were 58.8 and 89.1%, respectively. The decision curve analysis indicates a net benefit in prediction of severe AKI using the model across a range of threshold probabilities between 5 and 35% (Figure 4), illustrating the additional benefit of the model in predicting severe AKI over and above the approaches of “intervention for all” and “intervention for none”.

According to the PROBAST assessment, 19 of 20 signaling questions were rated as “Yes,” and 1 was rated as “Probably Yes,” thus the overall risk of bias of the final model was rated as low risk of bias. Detailed rationales of answers were described in Supplementary Table S3.

From the above computation, the following risk prediction equation was derived:

where P_outcome indicates the probability for the study outcome occurring, and A = −0.0101 ^* eGFR + 0.0161 ^* Heart Rate + 0.0378^*APACHE II score + 0.4577^* (if sepsis) + 0.3747^* (if admitted from hospital floor) + 0.2080^* (if admitted from OR after elective surgery) + 0.2577^* (if admitted from OR after emergency surgery) + 0.3699^* (if admitted from other hospitals) + 0.2154^* (if mechanical ventilation on admission) – 4.1529.

The prediction model was translated using REDCap to an online calculator for readers' convenience (to visit, please scan the QR code in Supplementary Figure S1 using mobile devices or visit: http://redcap.scrds.net/surveys/ and enter code: 7KMH8HN7N). The probability would be shown automatically at the completion of all predictors. Readers are encouraged to submit their data that can be collected for further improvement of the model.