Study design
This study was approved by the Peking University First Hospital Ethics Committee, and the requirement for written informed consent was waived.
Data source and Study Patients
This study used data obtained from the perioperative database of Peking University First Hospital, which contains the perioperative information of inpatients from 2012 onward. This study analyzed data from adults (age ≥ 18 years old) who underwent elective non-cardiac surgery between January 1, 2012, and December 31, 2017. Non-cardiac non-kidney surgery was identified based on the International Classification of Diseases and Procedures, Ninth Clinical Revision Revision volume 3 (ICD-9-v3). All surgeries other than those with their ICD codes listed were defined as "other." (Table S1 in the Supplementary Appendix)
Patients were excluded from the study based on the following criteria, those that underwent cardiac surgeries, kidney surgeries, obstetric surgeries, local infiltration anesthesia, and missing perioperative data. Also, patients with more than one operation within a year (including reopening of surgical cases) were excluded.
The patients’ preoperative serum cholesterol levels were obtained from the laboratory database and linked to the hospital’s perioperative database. Results were categorized according to the day on which patients’ were operated on. The mean value (three month time frame) was calculated for each patient with more than one preoperative result.
Study ENDPOINTS
The endpoint was any patient with AKI within 7 days in the hospital. This study used KDIGO as the criteria for AKI, which was defined by the patient’s postoperative serum creatinine increase to not less than 26.5 μmol/l within 48 hours, or 1.5 times from the baseline within 7 days after surgery, or initialization of blood dialysis. As the serum creatinine level fluctuates much postoperatively and could cause an inaccurate estimate of Glomerular Filtration Rate (eGFR), creatinine this study did not define AKI based on the GFR value or urine output.
Statistical analysis
According to previous studies, the incidence of postoperative AKI would be 1.1-17.9% in patients that elected to have surgery.1-9 We expected the patients with an abnormal serum cholesterol level to have an OR of 1.20 when compared with the normal serum cholesterol level. With significance set at 0.05 and the power set at 90%, the calculated sample size needed to compare two proportions was 3206 patients in each group.
For the comparative analysis, patients were divided into two groups according to the occurrence of postoperative AKI. Continuous variables with a normal distribution were compared using the Student t-test, and those with non-normal distribution were compared with the Mann-Whitney U-test . The Kolmogorov-Smirnov test was used to determine whether the data were normally distributed or not . Categorical variables were compared using the Chi-Square test or continuity corrected Chi-Square test. Rank variables were compared using the Kruskal–Wallis H-test.
Logistic regression was used to detect any association between the concentrations of HDL and AKI
A logistic regression model was constructed using the following formula:
AKI = HDL + confounders
Confounders (covariates) were the same in both the logistic regression and generalized additive models, except for the HDL levels. Confounders were assessed based on a priori knowledge and other studies.1-9,11 The following covariates were considered: sex, age, BMI, revised cardiac risk index grade, surgery duration, anesthesia type, cancer surgery, intraoperative blood transfusion, surgical complexity (Modified John Hopkins hospital criteria, MJHSC 12, Table-S5 in Appendix supplement 1) and preoperative serum albumin and serum creatinine, anesthesiologist’s experience, intraoperative dexmedetomidine and colloid use.
The propensity score weighting analysis
Due to the huge systematic differences, this study balanced the patients with preoperative HDL below or above 1.03 mmol/L (the widely accepted threshold for cardiovascular risk13) by propensity score weighting. Propensity score weighting is a method to diminish the effect of measured confounding factors and to get a less biased result in observational studies. In the present study, propensity score weights were calculated by using gradient boosted regression models,14,15 in which high or low preoperative HDL was the dependent variable, and vectors of the following (age, gender, body mass index, revised cardiac risk index, surgery duration, anesthesia type, cancer surgery, intraoperative blood transfusion, surgical complexity, type of surgery classified by site, anesthesiologist’s experience, severe intraoperative hypotension, preoperative coronary heart disease, arrhythmia, cerebral infarction, diabetes, chronic kidney disease, preoperative creatinine, cholesterol components, intraoperative dexmedetomidine and colloid use) were the independent variables. Compared to the inverse probability of exposure weighting method (IPEW),16 the propensity score weights calculated by the gradient boosted regression models do not need to consider co-linearity and often get better balancing performance.14,15 Using the weights for the originally observed cohorts could create two new cohorts with the number of patients differing from the original, and may not be an integer (each patient was multiplied by a specific weight defined by the gradient boosted regression models). The logistic analysis with the propensity score weighting could lead to less biased results, i.e., a quasi-randomized study.
Statistical Packages
All data management and statistical analysis were performed using the R programming language (v.3.5.2).