Prognostic Factors for Postoperative Complications. An Aggregate Protocol for 10 Observational Studies From the Danish TRIPLE-A Cohort of 1.2 Million Surgeries.

Advances in minimally invasive surgery and anaesthesia have enabled a growing proportion of procedures to be performed safely as ambulatory or short‐stay surgery, with high patient satisfaction and economic benefit [ 145 , 146 ]. Consequently, low‐ and intermediate‐risk procedures traditionally requiring inpatient admission are increasingly managed with same‐day or single overnight discharge, while procedure‐specific risk profiles remain important [ 147 ]. Unplanned readmission, reoperation, and admission to the ICU are recognised quality indicators of ambulatory surgery and are associated with worse outcomes [ 148 , 149 ]. Large population‐based studies report unplanned hospital re‐visit rates of approximately 4%–5% in the United States and around 1% in Denmark for predominantly low‐risk procedures, with higher rates observed for selected operations [ 149 , 150 , 151 ]. Mortality will be described in a separate protocol using time‐to‐event methods, including Kaplan–Meier analysis, and development of a dedicated prediction model. To identify prognostic factors for unplanned readmission, reoperation, and ICU‐admission following surgery. Unplanned readmission within 30 days postoperatively. Reoperation within 30 days postoperatively. Unplanned ICU‐admission within 30 days postoperatively, excluding patients directly transferred to ICU after surgery. Unplanned readmission within 30 days postoperatively. Reoperation within 30 days postoperatively. Unplanned ICU‐admission within 30 days postoperatively, excluding patients directly transferred to ICU after surgery. No additional criteria beyond the shared methodology. Surgical procedures with: Planned readmission. Planned ICU‐stay postoperatively or preoperative ICU stay. Preplanned reoperation, such as two‐stage cruciate ligament reconstruction. Planned readmission. Planned ICU‐stay postoperatively or preoperative ICU stay. Preplanned reoperation, such as two‐stage cruciate ligament reconstruction. Distinguishing planned from unplanned events using registry data is challenging, despite exclusion of procedures typically associated with planned care, and may lead to misclassification. To mitigate this, cases with potential ambiguity will be manually reviewed, although some residual misclassification is likely.

A.P.H.K., C.F. and M.H.O. conceived the study. A.P.H.K., A.P., J.L., E.I.Ø., N.H., L.H.L., S.B.R. and M.H.O. developed the statistical analysis plan and study methodology. A.P.H.K. drafted the first version of the main protocol. A.P.H.K., A.P., C.F., A.S., C.L.K., M.W., E.I.Ø., M.F., H.H.L. and N.H. contributed to writing the individual study sections. C.F. managed the references. M.H.O., A.P. and E.I.Ø. contributed to data processing and coding of variables. C.L.K. validated outcome definitions. All authors contributed to data acquisition or interpretation, critically revised the manuscript, and approved the final version.

The authors have nothing to report.

This protocol describes 10 retrospective cohort studies, each focusing on specific prognostic factors for a specific complication during or after anaesthesia and surgery (Table  1 ). Each of these studies shares the methodology described in this section, adhering to the PROGRESS recommendations [ 15 ]. The protocol is developed in accordance with The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) [ 16 ] and the design was informed by the domains of the Quality in Prognosis Studies (QUIPS) tool to reduce the risk of bias in prognostic factor research [ 17 ]. Study‐specific details, such as definition of primary outcome and choice of candidate prognostic factors, are unique for each study and are described separately. Overview of outcomes in study 1–10. All studies are based on EHR‐data from the TRIPLE‐A database, which includes all patients who underwent surgical procedures in public hospitals in the Capital and Zealand Regions of Denmark between 1 January 2017 and 14 November 2025 [ 14 ]. These regions cover approximately half of the total Danish population. The following approvals cover the central aim of TRIPLE‐A: to analyse prognostic factors for perioperative complications and develop prediction models for risk stratification based on these analyses. TRIPLE‐A was deemed exempt from ethical approval by the Regional Research Ethics Committee (F‐25031347, 14 April 2025). Data extraction was approved by Centre for Health, Capital Region, Team for Medical Records Research (R‐25029803, 16 May 2025). The project was subsequently approved by The Capital Region Legal Department (p‐2025‐19193, 23 May 2025). Data from TRIPLE‐A is pseudonymised and handled in compliance with data protection regulations. Variables in the TRIPLE‐A Database are based on pre‐processed structured electronic health record (EHR) data coded by clinicians and data scientists in collaboration. Variables are validated by cross‐referencing them in the EHR system with subsequent adjustments in iterations until the variables reflect EHR data. Only aggregated data will be presented in the publications. Specific descriptions of data pre‐processing and validation have previously been published [ 14 ]. All EHR‐related data are drawn from the TRIPLE‐A Database, while data on prescription redemption data, socioeconomic status and ethnicity are collected from Statistics Denmark and the Danish National Prescription Registry. The two datasets will be merged via Statistics Denmark, Research Services (Forskerordningen) [ 18 ]. For each study, the target population is defined a priori as a selection of surgical or anaesthesiologic procedures chosen to best address the specific research question. Each study's intended case mix is presented in Table  2 . Ensuring high‐quality and complete outcome classification within the selected population is highly emphasised. For example, in studies of AKI, inclusion is limited to procedures where both pre‐ and postoperative creatinine measurements are routinely obtained, allowing classification of all patients rather than only those undergoing clinically indicated testing. This approach minimises selection bias and reduces the risk that identified prognostic factors merely reflect underlying frailty or selective testing in higher‐risk patients. If, during data curation or analysis, certain procedures are found to have insufficient outcome classification, low sensitivity for outcome detection, or other limitations compromising their suitability, the surgical case mix may be adjusted accordingly in the final analysis and reported in the manuscript. Surgical case mix in individual studies. Note: Airway Study 6: Includes general anaesthesia with airway management across all procedures. Abbreviation: OB/GYN, obstetric or gynaecological surgical procedures. Patients aged ≥ 18 years, undergoing surgery in general and/or neuraxial/peripheral regional anaesthesia. The exclusion criteria below are applicable for all studies, except study 6, which assesses airway management‐related outcomes occurring before surgery and is therefore less influenced by the type of surgical procedure: Surgical cases that lack data on central time points needed to create the perioperative timeline (e.g., no information on time for arrival to the operating room or start of surgery). Surgical procedures presented with fewer than 100 cases in the database, because low case volumes preclude reliable estimation of procedure‐specific complication rates and increase the risk of unstable regression estimates due to sparse data. Secondary surgical procedures within the same admission (reoperations). Surgical cases that lack data on central time points needed to create the perioperative timeline (e.g., no information on time for arrival to the operating room or start of surgery). Surgical procedures presented with fewer than 100 cases in the database, because low case volumes preclude reliable estimation of procedure‐specific complication rates and increase the risk of unstable regression estimates due to sparse data. Secondary surgical procedures within the same admission (reoperations). Definitions and data sources are listed in Table  1 . All outcomes will be coded generically for the entire population. Reliance on ICD‐10 coding alone leads to underreporting of complications [ 19 ]. Therefore, most outcomes are ascertained using several complementary sources, including indicative medication administrations, blood sample results, radiology examinations, registration of subjective symptoms, and text‐mining of clinical notes. These sources are coded into composite rules for detecting the outcome. For example, a therapeutic dose of heparin combined with detection of the term ‘pulmonary embolism’ in free‐text clinical notes, without nearby negations, is used as a signal for potential cases. To optimise sensitivity and specificity, these composite rules will be iteratively validated and refined through cross‐referencing with the manual EHR review in a subset of cases. Opioids will be converted into intravenous morphine equivalents (IME) using previously described conversion factors [ 20 ]. While the primary aim is prognostic rather than causal, causal terminology is used to guide appropriate covariate adjustment. To ensure a common language, we here list definitions [ 21 ]: Prognostic factors: The variables being assessed for associations with the outcome. Outcome: The endpoint of interest. All outcomes described in this protocol are binary (yes/no). Confounder: A variable that causally influences both the prognostic factor and the outcome and should be adjusted for to reduce bias. Mediator: A variable on the causal pathway between prognostic factor and outcome, which should not be adjusted for when estimating total effects. Covariate: A variable associated with the outcome but not causally related to the prognostic factor. Covariates may be adjusted for to improve precision but do not address confounding. Collider: A variable that is causally influenced by both the prognostic factor and the outcome. Conditioning on a collider induces bias and should be avoided. Ancestor: A variable that causally affects other variables in the model but has no direct causal effect on the outcome. Ancestors should generally not be adjusted for unless they act as confounders. Descendant: A variable that is affected by the prognostic factor but does not causally affect the outcome. Adjustment should be avoided, as it may introduce bias or dilute the prognostic factor's effect. Moderator (effect modifier): A variable that modifies the strength or direction of the prognostic factor–outcome relationship. Moderators are handled via interaction terms and are adjusted for when estimating the prognostic factor's effects conditional on the moderator. For each individual study, we will make directed acyclic graphs (DAGs) to identify relevant prognostic factors, confounders, colliders and mediators [ 22 ]. The DAG charts will be performed in the individual studies including a priori detection of unmeasured confounding (example in Supporting Information  S1 ). Colliders and mediators will be excluded from the analyses. A broad list of candidate variables for each individual study was selected from a list of TRIPLE‐A and Statistics Denmark variables based on clinical experience and using existing literature on prognostic factors for each study outcome. The full candidate variable list consists of demographic factors, comorbidities, procedural details, anaesthesia details, vital parameters, perioperative medication and blood samples (Table  3 ). In each study, candidate prognostic factors have been chosen and plotted in Table  3 . The lists will be reduced in the separate studies based on DAGs to avoid redundancy and collinearity, including the identification of variables representing the same underlying construct (e.g., weight and height versus BMI). In this process, we will focus on retaining variables with high clinical utility and availability. This will also apply for more specific measurement choices—for example, for the blood pressure construct, choosing between lowest, highest, or mean values of systolic, diastolic, or mean arterial blood pressure during surgery, anaesthesia, or stay at the operating theatre. The final list of variables will be carried forward to the below described statistical analysis. Candidate variables in each prognostic factor study. Surgical procedure Adjusting variable Hospital Adjusting variable Year of surgery Adjusting variable Prolonged LOS—defined as an LOS greater than or equal to the 75th percentile (in days) for each operation Baseline characteristics will be presented using descriptive statistics. For continuous variables, means and standard deviations (SD) will be used to summarise data with normal distributions, and medians and interquartile range (IQR) for non‐normally distributed data. Categorical variables will be reported as counts and percentages. In accordance with PROGRESS recommendations for prognostic factor research, we will estimate the adjusted association between each candidate prognostic factor (from the reduced list described above) and the outcome using a multivariable logistic regression model [ 15 , 23 ]. For outcomes where death constitutes a competing event, one or more of the following approaches will be applied. Logistic regression may be performed in the subset of patients alive at the assessment timepoint. Alternatively, outcomes may be defined as composite endpoints (e.g., ‘death or…’) to minimise bias from censoring by death. In addition, individual studies may perform sensitivity analyses using Fine–Gray competing risk models to account explicitly for the presence of competing risks. To reduce overfitting arising from the number of candidate predictors relative to the number of outcome events, penalised regression using Least Absolute Shrinkage and Selection Operator (LASSO) will be considered in each study [ 24 ]. LASSO enables shrinkage of regression coefficients and data‐driven variable selection to improve model stability and predictive performance [ 25 ]. The degree of penalisation will be determined using cross‐validation. We will estimate the influence of each candidate predictor with two estimates. First, odds ratios (OR) with the corresponding 95% confidence intervals (CI) will be reported for each prognostic factor, representing adjusted associations after mutual adjustment in the final model. Second, the relative contribution of individual predictors will be quantified as the change in area under the receiver operating characteristic curve (Δ‐AUC) following sequential removal of each factor from the full model, with performance compared against the discrimination of the full model. p values < 0.05 will be considered statistically significant. Primary conclusions will be based on effect estimates and confidence intervals rather than p values to emphasise clinical meaningful differences instead of merely statistical significance. Statistical analyses will be performed using the latest available stable version of R (R Core Team, Vienna, Austria). Surgical procedure, calendar year, and hospital site will be examined across all studies, with the specific approach (adjustment, stratification, or inclusion as candidate predictors) determined for each study. The extent of missing data will be reported for each candidate prognostic factor. The primary analysis will use either complete cases or multiple imputation, as determined prior to association analyses by each study based on the pattern and proportion of missing data. Where relevant, the other analysis will be performed as a sensitivity analysis. For handling missing data, we plan to use SMCFCS (Substantive Model Compatible Fully Conditional Specification) as the primary imputation method. If computational demands prove prohibitive, MICE (Multiple Imputation by Chained Equations) will be used as a secondary approach. To mitigate potential biases inherent to observational EHR‐based studies, several measures are applied. Study populations are defined a priori to ensure high completeness of outcome classification. Key variables and outcomes are based on predefined composite algorithms and validated against the full EHR. Confounding will be addressed through DAG‐informed variable selection and multivariable adjustment for predefined factors. Model overfitting is reduced using penalised regression with cross‐validation. Missing data will be handled using appropriate imputation strategies with sensitivity analyses where relevant. Finally, robustness of findings will be evaluated through prespecified sensitivity analyses and alternative model specifications. We aim to describe patterns of co‐occurrence between postoperative complications. For each study, the prevalence of other complications will be reported among patients with and without the study‐specific complication (see Table  2 ). As multiple complications are more likely to occur during prolonged or complex admissions, these associations are not intended to reflect causal relationships and should be interpreted as descriptive only.

The authors have nothing to report.

This protocol describes the methodology behind studies addressing 18 perioperative complications in a large surgical cohort from the Eastern Region of Denmark. Complications after surgery and perioperative management are major contributors to morbidity, mortality, and expenditures in healthcare [ 9 ]. Various risk prediction models including the ACS NSQIP Surgical Risk Calculator [ 26 ], the Preoperative Score to Predict Postoperative Mortality (POSPOM) [ 27 ], the Revised Cardiac Risk Index (RCRI) [ 28 ], the Clinical Frailty Scale [ 29 ], and MySurgeryRisk [ 30 ], have already demonstrated high discriminatory ability for predicting postoperative complications and mortality. Identification of prognostic factors with contemporary data may help inform clinicians and suggest factors for future predictive models. The findings from the described studies may contribute to future development of risk stratification approaches for perioperative complications, such as the TRIPLE‐A Safeguard. Such approaches could potentially be integrated into the EHR to support risk assessment at key time points during the perioperative course. This perspective highlights a possible clinical application of the results, while the primary aim of the present protocol remains to identify prognostic factors. The major strengths include: (1) generalisability by including a large variety of surgical procedures in each study; (2) the large population which enables a high event‐per‐variable ratio; (3) manually validated variables that resemble clinical everyday practice, cross‐referenced with the EHR by clinicians. Furthermore, the surgical classifications are validated by surgeons from respective surgical specialties based on SKS procedure codes and booking names (e.g., laparoscopic cholecystectomy or total hip arthroplasty) with the intention of creating clinically relevant surgical procedures in comparison to only using the booking name or procedure codes itself. Limitations include: (1) a retrospective study design that, although data is prospectively entered into the EHR, carries (a) a risk of missing data registration and (b) a risk of non‐standardised data registration because data was not originally intended for research purposes; (2) the possibility that, while variables in the database were validated, there may be subgroups in which data behaves differently and was not identified during the validation process; (3) a risk of residual confounding from unassessed variables.

The authors have nothing to report.

Postoperative complications significantly reduce the quality of care, impede recovery, increase hospital costs, and adversely affect patients' long‐term quality of life [ 1 , 2 , 3 , 4 ]. Prevention and early detection of complications, including cardiovascular events, organ dysfunction, and infections, are significantly associated with a reduction in lengths of stay and readmission rates [ 5 , 6 , 7 , 8 ]. Data suggest that complications may account for 30% of surgical care costs [ 9 ]. The risk of complications varies across patients, surgical procedures, and perioperative care strategies, reflecting differences in underlying risk profiles [ 10 ]. Understanding the prognostic impact of these factors is necessary for the development of future prediction models that enable a priori or real‐time risk stratification [ 11 ]. Such models can support individualised planning of the perioperative stay and prehabilitation, both of which have been associated with improved survival and reduced healthcare costs [ 12 ]. Large electronic health record (EHR)‐based databases can provide relevant data granularity on patient characteristics, surgical details, anaesthesia management, vital parameters, and biomarkers, which allows a wide range of candidate prognostic factors to be assessed at an acceptable event‐per‐predictor ratio [ 13 ]. This protocol describes a common methodology and specific backgrounds for 10 prognostic factor study domains assessing 18 specific postoperative complications. The studies will be carried out using the AI and Automation in Anaesthesia (TRIPLE‐A) Database, which currently holds over 1.2 million surgical cases with a broad selection of validated clinical variables for each case across multiple domains [ 14 ]. We aim to assess prognostic factors for different short‐ and long‐term, intra‐ and postoperative complications in 10 domains: (1) acute postoperative pain, nausea/vomiting (PONV), and need for rescue opioids; (2) persistent opioid use; (3) acute kidney injury (AKI); (4) venous thromboembolism (VTE); (5) staphylococcal surgical site infection and infections requiring antibiotic treatment; (6) difficult airway management; (7) transfusion requirements; (8) delirium; (9) new‐onset postoperative atrial fibrillation; and (10) risk of readmission, reoperation and unplanned ICU‐admission.

The authors declare no conflicts of interest.

Supplemental S1: Example of a direct acyclic graph (DAG) illustrating hypothesised relationships between candidate prognostic factors and postoperative infection. Supplemental S2: ICD codes, free‐text terms, and ATC codes used for outcome identification.