Acute Kidney Injury Risk Prediction Score for Critically-ill Surgical Patients

Konlawij Trongtrakul; Jayanton Patumanond; Suneerat Kongsayreepong; Sunthiti Morakul; Tanyong Pipanmekaporn; Osaree Akaraborworn; Sujaree Poopipatpab


BMC Anesthesiol. 2020;20(140) 

In This Article


Study Design and the Source of Dataset

The prediction score was a secondary analysis from a prospective observational study, the Thai Surgical Intensive Care Unit (THAI-SICU) Study. It was conducted from 9-University base ICUs around Thailand between April 2010 to January 2013. Data was collected for 28 days following the ICU admission. A total of 4652 cases were collected with many outcomes of interest concerning complications following major non-cardiothoracic operations; for instance acute respiratory distress syndrome, delirium, and readmission, as which reported elsewhere.[10–12]

One topic of interest from the THAI-SICU Study was AKI outcome. In summary, the incidence of AKI, using the Acute Kidney Injury Network (AKIN) classification, was at 16.9%[1] and incidence remained high at 19.3% when specified only in the elderly group, whose age was equal to or over 65 years old.[12] In the total cohort, renal replacement therapy (RRT) was commenced in about one-fifth (22.3%) of AKI patients. AKI is associated with bad outcomes including greater ICU mortality and 28-day mortality. The risk factors for developing AKI included a higher severity of illness as measured by APACHE-II scores, the presence of hypoalbuminemia, and organ dysfunction from the start of ICU admission.


Critically-ill surgical patients from the THAI-SICU Study with aged 18 and over who underwent major non-cardiothoracic surgery before admission to ICU were eligible for enrolment into the study. The exclusion criterion were patients admitted to the ICU for less than 24 h or who were admitted to the ICU due to medical rather than surgical reasons; for instance, congestive heart failure, volume overload, or exacerbation of airway diseases that had no association or correspondence with surgical interventions.

Outcomes and Definition of AKI

The primary outcome was the presence of AKI within first 7 days of ICU admission. AKI was defined according to the KDIGO criteria,[13] which are an increase in serum creatinine (sCr) ≥ 0.3 mg/dL within 48 h or an increase of 1.5 times from baseline within a 7-day period.

The original dataset collected the incidence of AKI according to AKIN classification.[1] However, we extracted raw database that contained every single serum creatinine measurement during the ICU admission for reckoning AKI according to KDIGO criteria.

Reference sCr was selected according to the lowest sCr between the lowest value of sCr during ICU admission[14] or calculated back from MDRD equation by assuming patient's baseline estimated glomerular filtration rate (eGFR) at 75 mL/min.[15,16] In cases with a known history of chronic kidney disease, the best 3-month sCr preceding ICU admission was used as the reference value

Other secondary outcomes were also extracted: ICU mortality rate; day-28 mortality rate; ICU length of stay; and hospital length of stay.


Baseline characteristic data were utilized to deliver AKI prediction score including patient demographics (age, gender, body weight, and body mass index); pre-existing comorbidities (diabetes mellitus, hypertension, cardiovascular diseases, chronic pulmonary diseases, chronic kidney disease, malignancy, and others); severity of illness at ICU admission (measured by APACHE-II score, SOFA score, and SOFA non-renal score); sepsis at ICU admission; basic laboratory investigations at ICU admission (hemoglobin, serum albumin, blood sugar, PaO2/FiO2 ratio, chest imaging, electrocardiography, sCr, and reference sCr); and perioperative data before the ICU admission (including the American Society of Anesthesiologists ASA classification, emergency surgery, site of operation, perioperative time, blood loss, fluid balance, and urine output).

Sample Size

The effective sample size to enhance the statistical power in our study was calculated according to the most commonly mentioned "rule of thumb, 10 events needed per predictor".[17–19] There were almost 32 possible predictors included in the model. That meant, the event of AKI should be about 320 (32*10) cases. The original dataset had an incidence of AKI at 16.9%. The suitable sample size enrolled to develop the scoring system should be at least 1893 (100/16.9*320) cases, in which cases from ours (4652 cases) were enough to build the model.


Although we had tried hard to collect and clean data, missing values are inevitable. So, complete case analysis was used in our study.

IRB Committee and Consent, TCTR

The Institutional Review Board's approval for the study was obtained (Faculty of Medicine Vajira Hospital, Navamindradhiraj University, Bangkok, Thailand, COA 60/2561 and the Faculty of Medicine, Thammasat University, Pathumthani, Thailand, MTU-EC-ES-0-084/61), and internationally registered at, TCTR20190408004. Informed consents were waived by reason of a secondary analysis of the dataset.

Statistical Analysis

Categorical data were expressed with frequencies (n) and percentages (%) and compared using Fisher's exact test. Continuous data were presented in mean and standard deviation (SD) or median and interquartile range (IQR) and compared by Student's t-test or Wilcoxon's rank-sum test, as appropriate.

Model Development

To identify predictors that determined AKI, all predictors were first tested for multi-collinearity using variance inflation factor (VIF) > 10 criteria, and then entered into the model using multivariable logistic regression analysis. The possible significant variables were selected using criteria of p-value < 0.05 by backward elimination method. Categorization for continuous variables was done to facilitate odds ratio calculation.

Score Derivation and Validation

The prediction score for each independent variable was created by calculating its multivariable logistic regression coefficients divided by the lowest value of the model and rounded to the nearest integer or 0.5. Each predictor score was summed up to a total AKI prediction score. The final score was tested for its discriminative ability using an area under the receiver operating characteristic curve (AuROC) or C-statistic.[20] Scoring calibration between predicted risk and observed risk were compared and presented graphically, and were tested by the Hosmer-Lemeshow Goodness-of-fit (HL-GOF) statistic. Internal validity was done by the bootstrapping method (1000 replications). Finally, the prediction scores were categorized into four levels of AKI probability: low, moderate, high, and very high risk. A positive likelihood ratio (LH+) of AKI and its 95%CI were reported for each level.

All analyses were performed using STATA statistical software version 13.0 (StataCorp LP, College Station, TX, USA) and p-values of less than 0.05 were considered statistically significant.