This single-centre retrospective cohort study was approved by the Nagoya City University Graduate School of Medical Sciences and the Nagoya City University Hospital Institutional Review Board (60-21-0155, 15 March 2022). The manuscript adheres to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.17 A data analysis and statistical plan was written after the data were accessed.

Participants

We included adult patients (aged ≥ 18 yr) with an American Society of Anesthesiologists Physical Status (ASA-PS) I–IV who underwent noncardiac and nonobstetric surgery with duration of ≥ 60 min under general anesthesia in Nagoya City University Hospital between January 2015 and December 2021. Additional inclusion criteria were patients without mechanical circulatory support before and during the surgery, patients who experienced 5–120 min of anesthesia induction time, defined as the duration between induction and skin incision,5 and patients who underwent tracheal intubation or supraglottic airway device insertion at least once. After the automated extraction of eligible patients using the anesthesia information management system (AIMS), patients with incorrect data (ASA-PS classification or surgical type) or missing characteristic data were excluded by reviewing the electronic medical records (EMR) according to the following rules. First, patients who underwent surgery for intracranial hemorrhage with mass effect or aortic rupture were reclassified as ASA-PS V and excluded from the study.18 Second, patients who actually underwent cardiac or obstetric surgery were excluded. Finally, patients with missing data (characteristic data, blood pressure [BP] values, or postoperative outcomes) were excluded. If more than one surgical record met the inclusion criteria for the same patient, the most recent one was analyzed.

Data sources

We collected patient data from the EMR, an AIMS (Fortec ORSYS, Koninklijke Philips N.V., Amsterdam, Netherlands), and a diagnosis procedure code (DPC) database of our hospital. All data were obtained between June and December 2022. The DPC database is used for the medical fee billing of inpatients in Japan and has been used in several epidemiological studies.19,20,21 The DPC database includes administrative claims data and patient data, such as diagnoses and comorbidities at admissions identified by the International Classification of Diseases 10th Revision (ICD-10) codes, surgical procedures identified by the Japanese original codes (K codes), and malignancy.20 A previous validation study of the DPC database reported that malignancy has 83.5% sensitivity and 97.7% specificity.22

Variables

Patient characteristics, including age, sex, height, weight, body mass index (BMI), and ASA-PS class, were collected from the AIMS. Outliers in the patient height and weight data with > 3 standard deviations (SDs) were manually checked to determine if they were identical to the EMR data. If the outliers were erroneous and the correct values were obtained from the EMR, they were corrected. In addition, patients with a BMI of 30–40 kg m−2 and those with a BMI ≥ 40 kg m−2 were reclassified to be at least ASA-PS ≥ II and ≥ III, respectively.18 Anesthetic and surgical data, such as duration, emergency surgery, anesthesia type (total intravenous or inhalational), neuraxial anesthesia (epidural or spinal), and peripheral nerve block, were extracted from the EMR and AIMS. Furthermore, the K codes were extracted from the DPC database, and each surgery was classified for three-level risk scores, i.e., low, intermediate, and high risk, defined by the European Society of Cardiology and the European Society of Anaesthesiology.19,23

The most recent blood testing data within 90 days preoperatively were extracted from the EMR. Preoperative medication types were also extracted from the EMR. To calculate the Charlson’s comorbidity index, the ICD-10 codes for comorbidities and malignancies were extracted from the DPC database.24

Endpoints

The primary and secondary endpoints were the 30-day and 90-day mortality, respectively. These endpoints were extracted from the EMR. For patients discharged or transferred to another centre, the last outpatient visit data and mortality records written in the medical information forms from other centres were investigated.

Preprocessing of blood pressure data

Data preprocessing of BP is described in Electronic Supplementary Material (ESM) eAppendix. Hypotension was defined as a MAP below 65 mm Hg.1,2,5,6,7,8 We used absolute values because there was no benefit to using relative rather than absolute thresholds for postoperative organ injuries in a large retrospective study.8 Additionally, we used a threshold of 65 mm Hg based on a systemic review that showed an increased risk of organ injuries with MAP < 65 mm Hg.25 To quantitatively evaluate the degree of hypotension, the following four indices were calculated: duration, time proportion (hypotensive duration divided by the evaluated duration), area under the curve (AUC) of BP values < 65 mm Hg, and the proportion of AUC (AUC divided by the evaluated duration [anesthesia induction time or surgical duration]). The proportion of AUC was defined as the most important exposure because it included the severity and duration of hypotension and was time-adjusted.

Statistical analysis

Data are reported as number (proportion) for categorical data and median [interquartile range] for continuous data. For comparisons of two groups, we used Fisher’s exact test for binomial variables and the Mann–Whitney U test for continuous variables.

Multivariable logistic regression model

As a primary analysis, we used multivariable logistic regression models to reveal the relationship between hypotension and postoperative mortality. No sample size was calculated. The total number of patients in our cohort was expected to be approximately 16,000 cases. Based on recent reports, we assumed that the incidence rates of 30-day and 90-day mortality in our cohort would be 0.3%–0.5% and 0.8%–1.2%, respectively.1,26,27,28 Considering a 10% dropout rate of patients because of missing variables, 30-day and 90-day mortality events were estimated as 40–70 and 100–180, respectively. Considering the model stability, four explanatory variables for 30-day mortality and eight for 90-day mortality classification were included in the logistic regression models.29 From a clinical standpoint and previous literature, age, ASA-PS (dichotomized as ASA-PS < III and ≥ III), emergency, and surgical risk (dichotomized as low risk and others) were initially selected as explanatory variables for the 30-day mortality classification models.14,15 For the 90-day mortality models, age, sex, BMI, ASA-PS (four classes), emergency, surgical risk (three classes), anesthesia type, and neuraxial anesthesia were prespecified.

Sensitivity analyses

We performed sensitivity analyses using the PSM and RUSBoost models for the robustness of the results. Patients whose DPC data were unavailable or whose induction time was < ten minutes were excluded. Missing values of the blood tests were imputed by the mean values.

Propensity score matching model

We compared postoperative mortality rates between the two groups divided by the degree of PIH using PSM. For intuitive understanding and evidence of a significant increase in mortality in MAP < 65 mm Hg for ten minutes,25 patients were divided using the duration of hypotension, i.e., < 10 or ≥ 10 min of PIH. The propensity scores were calculated using 11 patients’ characteristics, 13 types of preoperative medication, 12 types of blood testing data, and Charlson’s comorbidity index score for each 30-day and 90-day mortality model. These 37 variables were determined based on clinical knowledge and previous literature. Propensity score matching was conducted with a 1:1 ratio by nearest-neighbour matching using 0.2 of the caliper width for the SD difference in the propensity scores without replacement.

RUSBoost model

We created machine learning models to predict postoperative mortality in each of the four hypotensive indices, and we evaluated the effects of hypotension using feature importance. As postoperative mortality rates were expected between 0.3% and 1.2% in our cohort, we used the RUSBoost model to classify imbalanced data.16 The RUSBoost algorithm consists of random undersampling and boosting algorithms. In these algorithms, the major class (survivor) is undersampled to a defined ratio to the minor class (deceased), and multiple models are trained sequentially to improve the classification performance. Feature importance measures how much a feature contributes to classifying a target and is used to interpret the relationship among variables.30 To construct RUSBoost models, patients were divided into training and test cohorts with a 7:3 ratio. Hyperparameters such as the learning rate and max depth were determined by GridSearch with five-fold cross-validation. Feature importance values were calculated for the patients’ characteristics and hypotensive indices of PIH and IOH. The performance of the created RUSBoost models for predicting postoperative mortality was assessed by the test data set using precision, recall, geometric mean (G-mean), and area under the precision–recall curve, considering the imbalanced nature of the outcome. The G-mean is the square root of the product of sensitivity (recall) and specificity. In total, 100 repeated trials were performed to precisely evaluate feature importance and model performance.

In addition, we performed sensitivity analyses using different definitions of hypotension and various subgroup analyses. All P values < 0.05 were considered statistically significant. Data preprocessing and model constructions were performed with Python 3.8.13 (Python Software Foundation, Wilmington, DA, USA) and scikit-learn 1.1.1,Footnote 1 and statistical analyses were performed with R version 4.0.5 (R Foundation for Statistical Computing, Vienna, Austria).