This study was a retrospective, single-center cohort analysis conducted at The Affiliated Traditional Chinese Medicine Hospital, Southwest Medical University, a tertiary referral hospital in Southwest China. The investigation covered the period from January 2018 to December 2023 and included adult patients who underwent elective major abdominal, urologic, or gynecologic surgery under general anesthesia with continuous invasive arterial pressure monitoring. Because of the retrospective nature of the analysis, no patient contact or intervention was involved. All perioperative data were extracted from pre-existing electronic systems, and therefore this study posed no additional risk to participants. The research protocol was approved by the institutional ethics committee (Approval No. BY2025088), which waived the requirement for informed consent due to the retrospective design and anonymization of data.

All adult patients (≥18 years) who met the inclusion criteria during the study period were consecutively enrolled to minimize selection bias. Inclusion criteria were: (1) elective major abdominal, urologic, or gynecologic surgery with invasive arterial blood pressure monitoring from induction to emergence; and (2) complete perioperative clinical information available in the anesthesia information system. Exclusion criteria included: (1) missing >5% of intraoperative arterial pressure data or >5 consecutive minutes of signal loss; (2) end-stage renal disease, advanced hepatic failure (Child–Pugh class C), or New York Heart Association class IV heart failure; (3) intraoperative death or aborted procedures; and (4) emergency surgery. All eligible cases were identified through a complete query of the anesthesia information management system (AIMS), ensuring comprehensive inclusion and eliminating sampling bias. No matching or prospective follow-up was performed because all data were derived from completed medical records.

The primary exposure was intraoperative hypotension, defined as mean arterial pressure (MAP) < 65 mmHg, measured continuously via invasive arterial line. Outcome variables included a composite of (1) acute kidney injury (AKI, defined per KDIGO criteria), (2) postoperative delirium (confirmed using CAM-ICU and DSM-5 criteria), (3) unplanned ICU admission within 48 h, or (4) all-cause 30-day mortality. Secondary outcomes comprised reoperation within 7 days, surgical site or organ-space infection (CDC/NHSN criteria), ICU readmission within 7 days, and postoperative hospital length of stay. For the purposes of this study, outcomes were ascertained over prespecified postoperative windows: AKI and postoperative delirium were evaluated from the end of surgery through postoperative day 7 or hospital discharge (whichever occurred first); unplanned ICU admission within 48 h, ICU readmission within 7 days, and reoperation within 7 days were identified from time-stamped ICU and operating room records; 30-day all-cause mortality was determined from EMR problem lists, discharge summaries, and the institutional death registry; and postoperative hospital length of stay was defined as the number of days from the index surgery to hospital discharge.

All perioperative information was obtained from the hospital’s electronic medical records (EMR) and AIMS, which automatically capture minute-by-minute invasive MAP values. Data acquisition hardware and software remained uniform throughout the study period. Automated artifact detection algorithms identified erroneous waveforms, which were manually reviewed by two investigators blinded to outcomes. Short gaps (<5 min) were linearly interpolated, while longer missing segments led to case exclusion. Laboratory and outcome variables were physician-verified at data entry. All postoperative outcomes—including AKI, delirium, ICU admissions and readmissions, reoperations, infections, mortality, and length of stay—were ascertained exclusively from the hospital EMR using a combination of structured diagnosis fields, laboratory values, ICU flow sheets, and narrative discharge summaries; these outcome variables are not recorded in the AIMS, which is limited to intraoperative monitoring and anesthesia-related data. Diagnostic definitions (KDIGO, DSM-5, CDC/NHSN) ensured consistency across patients. The EMR and AIMS datasets were merged using unique patient identifiers and surgery timestamps, and linkage accuracy was verified by cross-checking overlapping variables such as procedure type, date, and anesthesia start time; concordance for these linkage fields exceeded 99%, indicating reliable matching between the two systems rather than duplication of outcome information. All outcome data were extracted for research purposes only after at least 30 postoperative days had elapsed for the last included surgery, ensuring a complete 30-day follow-up window for every patient in the cohort.

To address potential bias inherent in retrospective analyses, multiple strategies were implemented. First, consecutive inclusion of all eligible surgeries during the study period minimized selection bias. Second, standardized data collection from automated monitoring systems reduced measurement bias. Third, independent double verification of critical variables (blood loss, operative duration, outcome classification) and adjudication by blinded reviewers minimized misclassification. Finally, statistical adjustment for key clinical confounders (age, ASA class, comorbidities, intraoperative factors) reduced residual confounding.

Because the study was retrospective and utilized existing electronic anesthesia records, all eligible cases during the five-year period were included. Nevertheless, the sample size was verified by a priori estimation using preliminary data, assuming a 20% baseline event rate and an odds ratio of 1.8 for major complications, which indicated that at least 720 cases were required to achieve 80% statistical power at α = 0.05. The final sample of 789 patients exceeded this threshold, ensuring robust analytic validity.

Continuous variables were summarized as mean ± standard deviation (SD) or median with interquartile range (IQR), depending on normality (assessed by Shapiro–Wilk test). Categorical variables were expressed as counts and percentages. For trajectory classification, hypotension duration and recurrence were grouped into clinically meaningful categories (<10 min, 10–30 min, and >30 min or ≥3 episodes) based on prior literature and the observed distribution within the dataset.

Statistical analyses were performed using R software (version 4.2.2; R Foundation for Statistical Computing, Vienna, Austria). Group-based trajectory modeling (GBTM) was applied to minute-by-minute mean arterial pressure (MAP) data using the traj package to identify latent intraoperative hypotension patterns. MAP was modeled as a function of absolute intraoperative time (minutes since induction) using third-order (cubic) polynomial terms of time within each latent class, assuming a censored normal distribution for the continuous MAP outcome. We initially fitted models with two to five trajectory classes and compared them using the Bayesian Information Criterion (BIC), mean posterior class-membership probabilities, minimum group size, and clinical interpretability of the resulting patterns. The final three-class solution was selected because it had the lowest BIC among clinically plausible models, mean posterior probabilities ≥0.80 for all classes, and no trajectory group comprising <10% of the cohort; procedures lasting longer than 180 min were truncated at 180 min, whereas shorter procedures contributed all available MAP observations without padding, as the GBTM framework accommodates trajectories of unequal length. For comparisons across the three trajectory groups, overall p-values in Table 1 were derived from global Chi-square tests for categorical outcomes (or Kruskal–Wallis tests for non-normally distributed continuous variables), whereas the p for trend values were calculated by treating the trajectory category as an ordinal variable (Trajectory A = 1, B = 2, C = 3) in logistic regression and testing for a monotonic increase in complication rates.

Independent predictors of the primary composite outcome were determined by multivariable logistic regression, including variables that were clinically relevant or had a univariable p < 0.10; multicollinearity was assessed using variance inflation factors, with VIF > 5 considered exclusionary. Model discrimination was assessed by the area under the receiver operating characteristic curve (AUC) and calibration by the Hosmer–Lemeshow goodness-of-fit test and calibration plots. To assess the incremental predictive value of hypotension trajectory grouping, three nested models were compared: Model 1 (clinical variables only), Model 2 (clinical variables plus trajectory group), and Model 3 (trajectory group only). In additional exploratory analyses, we also constructed logistic regression models in which conventional IOH exposure metrics—nadir MAP, cumulative minutes with MAP <65 mmHg, and the time-weighted average (TWA) MAP <65 mmHg—were entered as predictors, either alone or in place of the trajectory variable; the performance of these conventional-metric models is summarized in Supplementary Table S1 to facilitate direct comparison with the trajectory-based models. Internal validation was performed using bootstrap resampling (1,000 replicates) to adjust for optimism. Missing data (<5% for all variables) were imputed using multiple imputation with chained equations under the missing-at-random assumption. Decision curve analysis (DCA) and clinical impact curves (CIC) were used to quantify clinical utility and net benefit across risk thresholds. All tests were two-tailed, and p < 0.05 was considered statistically significant.

Baseline demographic and intraoperative characteristics of the study cohort, stratified by intraoperative hypotension trajectory patterns, are summarized in Table 2. Among the 789 patients, 253 (32.1%) were classified as Trajectory A (brief, mild hypotension), 308 (39.0%) as Trajectory B (moderate sustained hypotension), and 228 (28.9%) as Trajectory C (prolonged or fluctuating hypotension). As detailed in Table 2, baseline and intraoperative characteristics differed significantly across the three trajectory groups. In general, patients in Trajectory C were older, had more comorbidities, and experienced longer and more hemodynamically demanding operations, whereas Trajectory A included younger patients with fewer comorbidities. The temporal patterns of mean arterial pressure (MAP) for each trajectory group are depicted in Figure 1. Trajectory A exhibited stable MAP values with only transient, mild decreases below the hypotension threshold (65 mmHg). Trajectory B demonstrated sustained moderate hypotension lasting 10–30 min, while Trajectory C showed repeated or prolonged episodes of hypotension, often markedly below the threshold, reflecting greater intraoperative hemodynamic instability.

The incidence of postoperative complications differed significantly across intraoperative hypotension trajectory groups (Table 1). The primary composite complication occurred in 21.3% of patients overall, with a stepwise increase from 13.4% in Trajectory A to 20.8% in Trajectory B and 30.7% in Trajectory C (p < 0.001 for both overall comparison and trend). Among individual components, the rates of acute kidney injury (AKI), postoperative delirium, ICU admission within 48 h, and 30-day all-cause mortality all increased progressively from Trajectory A to C (all p for trend ≤ 0.004). Notably, AKI incidence was 18.0% in Trajectory C, more than threefold higher than in Trajectory A (5.9%). Secondary complications showed similar patterns. The frequency of reoperation within 7 days, surgical site or organ space infection, and ICU readmission within 7 days was highest in Trajectory C. Median postoperative hospital stay was significantly longer in Trajectory C (12 days, IQR 9–17) compared with Trajectories A and B (9 and 10 days, respectively; p < 0.001 for trend). Overall, a clear dose–response relationship was observed, with more severe or prolonged hypotension trajectories associated with higher complication rates and longer postoperative recovery.

Multivariable logistic regression analysis identified intraoperative hypotension trajectory as an independent predictor of the primary composite complication (Table 3). Compared with Trajectory A, patients in Trajectory B had a 58% higher adjusted odds of developing the primary composite complication (adjusted OR 1.58, 95% CI 1.03–2.43, p = 0.035), while those in Trajectory C had more than double the risk (adjusted OR 2.42, 95% CI 1.54–3.80, p < 0.001), with a significant dose–response relationship (p for trend < 0.001). Other independent predictors included older age (per 10-year increase: adjusted OR 1.19, 95% CI 1.04–1.37, p = 0.012), ASA classification III–IV (adjusted OR 1.61, 95% CI 1.14–2.29, p = 0.007), coronary artery disease (adjusted OR 1.77, 95% CI 1.10–2.85, p = 0.019), preoperative hemoglobin <120 g/L (adjusted OR 1.49, 95% CI 1.06–2.09, p = 0.022), intraoperative duration >180 min (adjusted OR 1.42, 95% CI 1.02–1.98, p = 0.039), and estimated blood loss >500 mL (adjusted OR 1.72, 95% CI 1.17–2.53, p = 0.006). Vasopressor use showed a non-significant association (p = 0.112). These findings indicate that more severe or prolonged intraoperative hypotension trajectories remain strong, independent risk factors for major postoperative complications, even after adjusting for patient comorbidities and intraoperative variables.

Intraoperative blood pressure characteristics differed markedly among the three hypotension trajectory groups (Table 4). Patients in Trajectory C had the lowest mean baseline MAP (89.4 ± 11.1 mmHg) and the lowest mean intraoperative nadir MAP (56.4 ± 7.1 mmHg), compared with higher values in Trajectories A and B (both p < 0.001). The number of hypotensive episodes increased stepwise from Trajectory A to C (median 1 vs. 3 vs. 6 episodes; p < 0.001), paralleled by longer cumulative hypotension duration (5 vs. 20 vs. 45 min; p < 0.001) and longer single-episode durations. The time-weighted average MAP below 65 mmHg was markedly elevated in Trajectory C (5.3 ± 2.1 mmHg) compared with Trajectory B (2.6 ± 1.3 mmHg) and Trajectory A (0.9 ± 0.6 mmHg; p < 0.001), indicating more prolonged and severe hypotensive exposure. Vasopressor bolus requirements followed a similar gradient, with the highest median doses in Trajectory C [3 (IQR 2–5)] and the lowest in Trajectory A [1 (IQR 0–2); p < 0.001]. These findings confirm that the predefined trajectory groups capture distinct patterns of intraoperative hemodynamic instability, with Trajectory C representing the most severe and sustained hypotension profile.

The comparative performance metrics of the three predictive models for primary composite complications are summarized in Table 5 and illustrated in Figure 2. Discrimination, assessed by the area under the ROC curve (AUC), was highest for Model 2 (Clinical + Trajectory) with an AUC of 0.860 (95% CI, 0.832–0.888), followed by Model 3 (Trajectory Only) at 0.834 (95% CI, 0.803–0.862), and lowest for Model 1 (Clinical Only) at 0.578 (95% CI, 0.531–0.624) (Figure 2A). Sensitivity and specificity were balanced in Models 2 and 3, with Model 2 achieving the highest sensitivity (77.0%) and a specificity of 76.2%, while Model 1 showed lower sensitivity (64.2%) and markedly reduced specificity (49.4%). Negative predictive values were highest for Model 2 (90.5%) and Model 3 (88.2%), indicating superior ability to rule out complications compared with Model 1 (79.8%). Calibration analysis of Model 2 (Figure 2B) demonstrated close alignment between predicted and observed probabilities, with both apparent and bias-corrected curves closely following the ideal line. Goodness-of-fit was supported by non-significant Hosmer–Lemeshow p-values for all models (p > 0.05), and Brier scores were lowest for Model 2 (0.173), reflecting superior overall accuracy. Decision curve analysis (Figure 2C) revealed that Model 2 provided the greatest net clinical benefit across a broad range of risk thresholds, consistently outperforming both Model 1 and Model 3 as well as the “treat all” and “treat none” strategies. To enhance clarity, the interpretation of each curve in Figure 2C has been elaborated in the legend. In particular, the magenta line corresponds to the combined model, the blue and green lines represent the trajectory-only and clinical-only models, respectively, and the black and gray dashed lines denote treat-all and treat-none strategies. This clarification allows the reader to more easily distinguish the net benefit patterns without altering the original figure. Clinical impact curves (Figure 2D) for Model 2 further illustrated its practical utility, showing a higher number of true positives among patients classified as high risk, particularly at clinically relevant thresholds. Collectively, these findings confirm that incorporating intraoperative hypotension trajectory patterns into a clinical model substantially improves discrimination, calibration, and potential clinical usefulness compared with a model based solely on clinical variables.

To complement these primary models, we also fitted exploratory logistic regression models incorporating conventional IOH exposure metrics (nadir MAP, cumulative minutes with MAP <65 mmHg, and TWA MAP <65 mmHg) in place of the trajectory classification. Their discrimination, calibration, and decision-analytic characteristics are reported in Supplementary Table S1, allowing readers to directly compare traditional IOH summary measures with trajectory-based classification.