Predicting postoperative acute kidney injury after colorectal cancer surgery: development and internal validation of a clinical prediction model using a nationwide cohort Authors: Mads Skovsgaard, Philip V Munch, Uffe Heide-Jørgensen, Lene H Iversen, Christian F Christiansen Published: 2026 ABSTRACT AbstractBackgroundPostoperative acute kidney injury (PO-AKI) occurs in 8–20% of patients undergoing colorectal cancer surgery and is associated with increased mortality and morbidity. A prediction model can help in the identification of high-risk patients and guide clinicians in implementing preventive measures. The present study aimed to develop and validate internally a clinical prediction model for the risk of PO-AKI within 7 days after an elective resection of colorectal cancer, using predictors known before surgery.MethodsUsing a nationwide clinical quality database, all patients with stage I–IV colorectal cancer undergoing an elective colorectal resection from 2015 to 2021 in Denmark were identified. Patient characteristics, comorbidities, medications, and laboratory values were obtained through linkage across national health registries. The outcome was PO-AKI within 7 days after surgery, defined according to the Kidney Disease Improving Global Outcomes criteria. A logistic regression model was fitted using preoperative predictors, backward selection was applied to derive the final model, and internal validation was performed through two-fold bootstrapping. The performance was assessed using discrimination and calibration measures.ResultsAmong 26 104 identified patients, 20 277 patients were included, and 1747 (8.6%) developed AKI within 7 days. Fifty-four per cent were men, and the median age was 71 years (interquartile range, 64–78 years). The final model included 14 preoperative predictors. The model showed moderate discriminative performance with an optimism-corrected c-statistic of 0.70 (95% confidence interval 0.68 to 0.71) and was well calibrated at PO-AKI risks between 0% and 20% but somewhat overestimated higher risks (overall slope 0.95; intercept −0.001). The prediction model is available as a calculator at https://po-aki.streamlit.app.ConclusionThe model predicted PO-AKI with moderate accuracy and may support the implementation of preventive measures. However, external validation is required to confirm its applicability.Keywords: postoperative complications, acute kidney injury, preoperative risk assessment, clinical decision support Introduction Acute kidney injury (AKI) is a multifactorial syndrome characterized by an abrupt decrease in kidney function identified by an increase in plasma creatinine or a decrease in urine output, as defined in the Kidney Disease: Improving Global Outcomes (KDIGO) guidelines1. Postoperative AKI (PO-AKI) is a frequent complication of major surgery2–4. Contributing factors include haemodynamic instability due to blood loss and fluid shifts, which reduces renal perfusion. Furthermore, the use of nephrotoxic agents (such as anaesthetics and antibiotics), inflammation, and oxidative stress from surgery can further impair renal function5. PO-AKI is associated with increased mortality and morbidity, and higher hospital costs, and even mild cases of AKI are associated with adverse outcomes3,6,7. For colorectal cancer (CRC), surgery is the cornerstone of treatment with curative intent, with a reported risk of PO-AKI ranging from 8% to 20%8–10. There are limited treatment options for AKI, making prevention crucial to improve outcomes. Preventive measures include haemodynamic optimization, close monitoring, and avoiding nephrotoxic medications1. Many models have been proposed to predict the risk of AKI after surgery, with most focusing on cardiac surgery11. Only one study12 has examined a model specifically for CRC surgery, and it only included 389 patients, limiting its generalizability and clinical applicability. This research therefore aimed to develop a prediction model with information known before CRC surgery, to aid targeted prevention in the preoperative phase, and to evaluate its potential clinical value through net benefit analysis. Methods The present study was designed for the development and validation of a prediction model and was reported according to the updated transparent reporting of a multivariable prediction model for individual prognosis (TRIPOD + AI)13. Data sources Patients undergoing CRC surgery were identified from the Danish Colorectal Cancer Group (DCCG) database. DCCG14 is a nationwide quality database that contains clinical information on all adult patients (age ≥ 18 years) diagnosed with primary CRC in a Danish hospital department since 2001. Using the personal identification number assigned to all Danish citizens at birth or immigration, DCCG data were linked with several national registries to obtain information on patient characteristics, comorbidities, medications, laboratory values, and vital status. Data on comorbidities were retrieved from the Danish National Patient Registry (DNPR)15, which contains information on all non-psychiatric hospitalizations since 1977 and all emergency room and outpatient visits since 1994. Information on redeemed prescriptions was retrieved from the Danish National Prescription Registry (NPR)16, which contains complete records of all prescriptions dispensed in Danish pharmacies since 1995. Laboratory measurements were obtained from the Laboratory Information Systems (LABKA) and the Register of Laboratory Results for Research (RLRR)17, which include results of blood tests from hospitals and general practitioners across Denmark, with complete nationwide coverage achieved in 2015. Finally, data on sex, age, and death were collected from the Danish Civil Registration System18, which has been continuously updated for changes in vital status and migration for the entire Danish population since 1968. Under Danish law, no ethical committee approval is required for registry-based research. The study was reported to the Danish Data Protection Agency through registration at Aarhus University (AU record No. 2016-051-000001/2624) and approved by Statistics Denmark (project No. 708570). This study was not preregistered in an independent institutional registry. Participants All patients who underwent a resection of the colon or the rectum for all stages of CRC from 1 January 2015, to 31 December 2021 from the DCCG database were included. The study group was restricted to patients undergoing resections to focus on major surgical procedures, which have higher risks of postoperative complications, and to achieve a greater homogeneity among the included patients. Table S1 details which codes were used to identify resections of either the colon and/or the rectum. Furthermore, the study was restricted to elective surgeries to ensure the model could be used in the preoperative phase, such as in a multidisciplinary conference setting. Patients were followed from the day of surgery until PO-AKI, death, or 7 days after surgery—whichever occurred first. To assess the baseline creatinine level, only patients with at least one plasma creatinine measurement within 365 days before surgery in the development of the prediction model were included. The baseline creatinine was defined as the most recent preoperative creatinine measurement within 365 days before surgery, and the baseline estimated glomerular filtration rate (eGFR) was calculated using the 2021 Chronic Kidney Disease Epidemiology Collaboration’s equation based on creatinine19. These measurements were obtained from both inpatient and outpatient settings. Patients with kidney failure at the time of surgery were excluded from the analysis, that is, patients who underwent dialysis within 365 days before surgery, had a previous kidney transplant, or had a baseline eGFR less than 15 ml per min per 1.73 m2. Finally, patients with AKI during the 30 days before surgery were excluded. Outcome of interest The outcome of interest was PO-AKI, defined in accordance with the joint consensus report of the Acute Disease Quality Initiative and PeriOperative Quality Initiative5 as AKI within 7 days after surgery. Here, AKI is defined using the 2012 KDIGO guidelines as either an absolute increase in creatinine of 26.5 µmol/l within 48 hours or a 1.5-fold increase in plasma creatinine compared with the most recent preoperative creatinine measurement. The creatinine measurements were obtained from the laboratory databases, LABKA and RLRR17. In this definition of AKI, decreased urine output was excluded, as urine output is not regularly measured outside intensive care units. Candidate predictors For the purpose of this analysis, a list of predictors that are routinely collected and obtainable in the preoperative phase to facilitate the implementation of this model before surgery was drawn up based on the literature and clinical knowledge of predictors for PO-AKI. The 19 candidate predictors and their definitions, including lookback periods and data sources, are detailed in Table S2. Data on age and sex, laboratory values for creatinine and haemoglobin, performance status, comorbidities, use of medications, information on cancer stage, intended operative approach, and the intention of the surgical procedure (provided by the DCCG database) were collected. The included comorbidities were diabetes mellitus, chronic kidney disease (CKD), heart failure, ischaemic heart disease, liver disease, chronic obstructive pulmonary disease (COPD), and hypertension. These comorbidities were defined mainly using ICD-10 codes identified in the DNPR from in-hospital contacts, emergency room visits, and outpatient specialist clinic visits for the condition within 10 years before CRC surgery15. CKD was defined as two outpatient creatinine tests separated by ≥ 90 days with an eGFR of < 60 ml per min per 1.73 m2 and identified using creatinine measurements from the laboratory databases. Furthermore, diabetes mellitus and hypertension were defined using a combination of the diagnosis codes and redeemed prescriptions of medications within 90 days before surgery in the NPR, as well as haemoglobin A1c measurements for diabetes mellitus (Table S2)16. Statistical analysis A sample size calculation is presented in the supplementary methods. It indicated that a minimum of 6440 patients were required to develop a robust prediction model. Baseline characteristics were summarized using medians with interquartile ranges (i.q.r.) for continuous variables and as frequencies and percentages for categorical variables. Baseline characteristics were reported both stratified by the outcome PO-AKI and for the overall study population. All analyses were performed using R statistical software20. Missing data were handled using multiple imputation with chained equations21; further details are provided in the supplementary methods. To develop the prediction model, multivariable logistic regression was applied to predict the risk of PO-AKI after elective resectional CRC surgery. The continuous covariates (age, body mass index (BMI), preoperative eGFR, and preoperative haemoglobin) were modelled with restricted cubic splines with five knots, placed at the 5%, 27.5%, 50%, 72.5%, and 95% percentiles, using the rcs function in the R package rms22. Backward elimination was used to identify variables for the final model, retaining those that reduced the Akaike Information Criterion (corresponding to a significance level of P = 0.157 for covariates with one degree of freedom). Predictors that were retained in at least eight (80%) of the models across the imputed data sets were included in the final model. The backward selection was carried out using the fastbw command from the R package rms22. Internal validation of the final model was performed using a two-stage bootstrap resampling process. This was done to obtain optimism-adjusted measures of model performance in the underlying population23. A detailed description of the internal validation process is provided in the supplementary methods. Model fit was evaluated using the Brier score, and discrimination was assessed with the c-statistic. For calibration, calibration-at-large, calibration slope, and intercept were assessed24. Calibration describes how the model performs across different risk groups, and calibration curves were constructed using 200 bootstrap resamples for each of the 10 imputed data sets (2000 bootstrap resamples in total). As part of the model evaluation, a net benefit analysis was also conducted, where using the model to guide treatment was compared with not taking any action or acting for everyone—regardless of AKI risk. This net benefit analysis was carried out to quantify the benefit of implementing the model in clinical practice, and is reported in a decision curve, where the net benefit is plotted against a range of cut-off thresholds above which closer monitoring or intervention would be considered25. Decision curve analysis was carried out using the dcurves package in R26. Results Participants A total of 26 104 patients undergoing surgery for CRC in the study period were identified in the DCCG database, of whom 20 277 patients met the inclusion criteria (Fig. 1). Among included individuals, the median age was 71 years (i.q.r. 64–78 years), 11 026 (54.4%) were men, and 1747 (8.6%) developed PO-AKI. Patient characteristics and the distribution of missing data in the cohort are presented in Table 1. Patients with PO-AKI were more often men, had a slightly higher BMI, and had more of the included comorbidities compared with patients without PO-AKI. The risk of PO-AKI remained relatively stable during the study period (Fig. S1). Using backward elimination, 14 predictors were retained in more than 80% of the 10 imputed data sets (all 14 were retained in all the data sets) and thus included in the final model. These were age, sex, BMI, World Health Organization performance status, hypertension, CKD, COPD, diabetes mellitus, liver disease, use of angiotensin-converting enzyme inhibitors or angiotensin-II receptor blockers, use of diuretics, preoperative eGFR, preoperative haemoglobin, and operative approach. Fig. S2 shows partial effect plots for the modelled restricted cubic splines, and the full equation for the model is reported in Appendix S1. Patient flow diagram illustrating identification of final study cohort Of 26 104 patients undergoing colorectal cancer surgery 2015–2021, a total of 20 277 were included for analysis after applying the specified exclusion criteria. Table 1Baseline characteristics of participants used to develop a model for the risk of postoperative acute kidney injury after elective resectional colorectal cancer surgeryOverall cohort (n = 20 277)No PO-AKI (n = 18 530; 91.4%)PO-AKI(n = 1747; 8.6%)Age (years), median (i.q.r.)71 (64–78)71 (64–77)74 (67–80)Sex ratio (male : female)11 026 : 92519864 : 86661162 : 585Body mass index (kg/m2), median (i.q.r.)*25.7 (23.1–29.1)25.7 (23.0–29.0)26.8 (23.6–30.6) Missing 280 (1.4%) 241 (1.3%) 39 (2.2%) WHO performance status at diagnosis Performance status 013 576 (67.0%)12 663 (68.3%)913 (52.3%) Performance status 14690 (23.1%)4131 (22.3%)559 (32.0%) Performance status 21217 (6.0%)1036 (5.6%)181 (10.4%) Performance status 3254 (1.3%)209 (1.1%)45 (2.6%) Performance status 439 (0.2%)32 (0.2%)7 (0.4%) Missing 501 (2.5%) 459 (2.5%) 42 (2.4%) Clinical UICC stage† 1(22%)(23%)(20%) 2(14%)(14%)(16%) 3(31%)(31%)(30%) 4(10%)(10%)(10%) Stage not evaluated(23%)(23%)(24%) Missing < 5 (0.0%) < 5 (0.0%) < 5 (0.0%) Medical history Hypertension5336 (26.3%)4679 (25.3%)657 (37.6%) Chronic kidney disease3808 (18.8%)3231 (17.4%)577 (33.0%) Chronic obstructive pulmonary disease719 (3.5%)610 (3.3%)109 (6.2%) Diabetes mellitus2767 (13.6%)2383 (12.9%)384 (22.0%) Liver disease210 (1.0%)182 (1.0%)28 (1.6%) Heart failure414 (2.0%)341 (1.8%)73 (4.2%) Ischaemic heart disease1771 (8.7%)1558 (8.4%)213 (12.2%) Medications within 90 days ACE inhibitors or angiotensin receptor blockers5811 (28.7%)5128 (27.7%)683 (39.1%) Diuretics1610 (7.9%)1341 (7.2%)269 (15.4%) Non-steroidal anti-inflammatory drugs1403 (6.9%)1293 (7.0%)110 (6.3%) Preoperative estimated glomerular filtration rate (ml per min per 1.73 m2), median (i.q.r.)85 (73–92)85 (74–92)81 (63–89) Preoperative haemoglobin (mmol/l), median (i.q.r.)8.1 (7.0–8.9)8.2 (7.1–8.9)7.8 (6.6–8.7) Missing 1628 (8.0%) 1527 (8.2%) 101 (5.8%) Operative approach† Open2610 (12.9%)(12%)(20%) Laparoscopic13 284 (65.5%)(66%)(59%) Robot assisted3985 (20%)(20%)(18%) Transanal total mesorectal excision386 (1.9%)(1.9%)(2.4%) Missing 12 (0.1%) (0%) (0%) Intent of surgery Curative19 832 (97.8%)18 131 (97.8%)1701 (97.4%) Palliative262 (1.3%)236 (1.3%)26 (1.5%) Compromised resection147 (0.7%)133 (0.7%)14 (0.8%) Missing 36 (0.2%) 30 (0.2%) 6 (0.3%) Values are n (%) unless otherwise stated. A total of 2223 included patients (11%) had missing data in at least one variable. *Unrealistic height and weight values, likely due to typing errors, were converted to missing values before imputing the data sets and computing the body mass indexes. Specifically, height values below 120 cm or above 250 cm, and weight values below 25 kg or above 500 kg were marked as missing. †In accordance with the rules for data confidentiality from Statistics Denmark, where the data were stored, some values were masked. To protect individual privacy, cells with fewer than five observations, or combinations of cells that could identify groups of four or fewer individuals, are not reported. PO-AKI, postoperative acute kidney injury; i.q.r., interquartile range, WHO, World Health Organization; UICC, Union for International Cancer Control; ACE, angiotensin-converting enzyme. Baseline characteristics of participants used to develop a model for the risk of postoperative acute kidney injury after elective resectional colorectal cancer surgery Values are n (%) unless otherwise stated. A total of 2223 included patients (11%) had missing data in at least one variable. *Unrealistic height and weight values, likely due to typing errors, were converted to missing values before imputing the data sets and computing the body mass indexes. Specifically, height values below 120 cm or above 250 cm, and weight values below 25 kg or above 500 kg were marked as missing. †In accordance with the rules for data confidentiality from Statistics Denmark, where the data were stored, some values were masked. To protect individual privacy, cells with fewer than five observations, or combinations of cells that could identify groups of four or fewer individuals, are not reported. PO-AKI, postoperative acute kidney injury; i.q.r., interquartile range, WHO, World Health Organization; UICC, Union for International Cancer Control; ACE, angiotensin-converting enzyme. Model performance The Brier score was 0.08 (95% confidence interval (c.i.) 0.07 to 0.08) and the optimism-corrected C-statistic was 0.70 (95% c.i. 0.68 to 0.71), implying a moderate discriminative ability. For the calibration, the mean predicted risk (calibration in the large) was 8.6%, the slope was 0.95 (95% c.i. 0.95 to 0.96), and the intercept −0.001 (95% c.i. −0.0008 to −0.0011). The model was well calibrated for risk predictions ranging from 0% to 20% but tended to overestimate the risk for probabilities exceeding 20% (Fig. 2). In the study population, 5.5% had a predicted risk above 20% (Table S3). Inspection of the decision curve (Fig. 3) showed that using the developed prediction model would result in higher clinical utility compared with assuming that all patients are at high risk or assuming that no one is at high risk. Using the model resulted in a higher net benefit than the treat-all and treat-none (where treating corresponds to initiating preventive measures) strategies for cut-off thresholds up to 30%, and similar to taking action for none above this. For instance, using a threshold of ≥ 10% risk of developing PO-AKI to determine patients at high risk results in a net benefit of approximately 0.017 (corresponding to a benefit of 1.7 true positives per 100 patients). Optimism-corrected calibration of prediction model and distribution of predicted probabilities The smoothed relationship between observed and predicted probabilities is depicted for the 2000 bootstrap samples with optimism correction. These relationships for the individual bootstraps are shown as thin grey lines, with the mean relationship shown in blue. Below the calibration curve, the distribution of predicted probabilities across the 2000 bootstraps is shown in a histogram. PO-AKI, postoperative acute kidney injury. Decision curve analysis for prediction of postoperative acute kidney injury The X-axis represents the threshold probability, which is the probability at which a patient would be considered high risk and thus receive a particular intervention. The Y-axis shows the net benefit of using the prediction model to guide clinical decisions. Net benefit is a measure of whether the model correctly identifies patients at high and low risk, where a net benefit of 8% corresponds to 8 true positives per 100 patients in the target population. The prediction model is available as a calculator at https://po-aki.streamlit.app. Discussion A prediction model for PO-AKI after CRC surgery was developed using variables available in the preoperative phase. The model demonstrated moderate discriminative performance and was well calibrated for predictions in the low-to-moderate range. However, it tended to overestimate higher risks. This suggests that, whereas the model performs reliably within the lower-risk range, predictions outside this range are less reliable, and adjustments may be needed to improve its accuracy for higher-risk predictions. Decision curve analysis indicated that the model offers clinical benefit over treat-all and treat-none strategies across a range of clinically relevant thresholds, supporting its potential utility in guiding preventive interventions. The risk of PO-AKI was 8.6%, which aligns with rates reported elsewhere, but differs from the previously reported 20.3% in Northern Denmark from 2005 to 20118. This discrepancy may be due to the previous study’s inclusion of emergency surgery, as well as changes in prehabilitation, surgical approach, AKI definitions, and coding practices over time. Fourteen of the 19 candidate predictors were included after backward elimination. Notably, all retained predictors—except for intended operative approach—were patient-related factors. Tumour stage was excluded from the model, possibly because it did not contribute to the prediction of PO-AKI after inclusion of other preoperative factors, such as performance status and comorbidities. Among surgical approaches, laparoscopic surgery was associated with the lowest predicted risk, even compared with robotic surgery. This may reflect the predominant use of robotic techniques in more complex rectal resections rather than colonic procedures. Although colonic and rectal resections differ in technical complexity, recent nationwide Danish data have demonstrated comparable 7-day risks of AKI following colectomy and rectal excision, suggesting that tumour location alone may not be a major determinant of early postoperative AKI risk27. Any potential differences are likely mediated through patient characteristics and perioperative factors, several of which are already captured in the present model. The model used clinical and laboratory values that are available before surgery and could therefore be implemented in the preoperative phase in clinical practice. The model was developed using a large nationwide cohort and is to the authors’ knowledge the first clinical prediction model for the risk of PO-AKI after CRC surgery based solely on preoperative variables. A previous model12 also aimed to predict PO-AKI in patients undergoing CRC surgery. It was developed using only 389 patients from a single centre and relied on a combination of preoperative, intraoperative, and postoperative predictors. In contrast, the authors’ model was developed using a large, nationwide cohort, using only variables available before surgery, and was internally validated with both calibration and decision curve analysis. These features enhance its applicability in preoperative decision-making. Other prediction models for PO-AKI in broader surgical populations have been developed, including the Simple Postoperative AKI Risk (SPARK)28 and Outcomes After Kidney injury in Surgery (OAKS)29 models in gastrointestinal surgery. The SPARK model, developed in over 90 000 patients undergoing non-cardiac surgery, demonstrated moderate discrimination (c-statistic 0.72) but showed highly variable calibration across surgical departments. The OAKS model, developed for gastrointestinal surgery and validated in the IMAGINE cohort of elective colonic surgery for ileus, showed lower discrimination (c-statistic 0.66) and tended to underestimate risk above 15%. Whereas both models are valuable tools with broad applicability, their generalizability may come at the expense of precision in more specific clinical settings. This model focuses on a more homogeneous cohort of patients undergoing elective CRC surgery. Although external validation is needed, it showed comparable discrimination, good calibration in the clinically relevant risk range, and uses only routinely available preoperative variables. These characteristics support its potential for individualized risk estimation and informed preoperative decision-making in this patient population. The model was both developed and validated using Danish registry and clinical data. To support generalizability, external validation is recommended before broader implementation in other settings. This is essential to confirm the model’s reliability and applicability across different populations. This is especially important as it was developed in a high-income setting in a predominantly White population. Additionally, model performance should be monitored over time, and model updating (for example recalibration) should be considered to reflect the differences in clinical and patient populations30. The model tended to overestimate the risk in higher-risk patients, which underscores the need for cautious interpretation in this range. However, this miscalibration would be inconsequential in practice if a lower threshold is used to define high/low risk. Similar miscalibration at higher predicted risks has been seen in other models developed for AKI31. Perioperative factors such as diverting stoma formation may influence the risk of postoperative renal impairment32. Although stoma formation is often planned before surgery, the final decision may change during surgery, and information on preoperative stoma intention was not available in the authors’ data sources. Consequently, stoma formation was not included, consistent with the aim to develop a prediction model based exclusively on variables reliably known before surgery. Another potential weakness is the use of creatinine measurements to assess PO-AKI. In many cases of uncomplicated CRC surgery, patients are discharged within a few days, and creatinine is thus not routinely checked in the follow-up period of 7 days. This may lead to an underestimation of the risk of PO-AKI, but it is suspected that these patients represent a lower-risk group and that missed AKI episodes will be milder cases and less clinically relevant. It is important to highlight that patients undergoing emergency surgery were excluded; although it would have been clinically relevant to study PO-AKI in emergency procedures, where the risk of PO-AKI is likely higher, this lies outside the scope of the present study. The model was developed specifically to support preoperative risk assessment, which may not be feasible in emergency surgical settings. By estimating the risk of PO-AKI, the clinical prediction model may help clinicians identify patients who may need extra attention before or during surgery. This can enable targeted preventive care, potentially improving outcomes. The prevention bundle in the 2012 KDIGO guideline includes volume management, maintaining adequate blood pressure and normoglycemia, and discontinuation or avoidance of nephrotoxic medications1. A liberal perioperative intravenous fluid regimen has been shown to reduce the risk of AKI within 30 days in a large, randomized trial of patients undergoing major gastrointestinal surgery33. For patients at higher risk, clinicians can adjust perioperative management, which may potentially help reduce the risk of PO-AKI. In summary, the authors developed a clinical prediction model with 14 variables that could predict PO-AKI with moderate accuracy. However, its predictive performance capability should be further validated in additional studies to ensure its reliability. Supplementary Material Contributor Information Mads Skovsgaard, Department of Clinical Epidemiology, Aarhus University Hospital, Aarhus, Denmark; Department of Clinical Medicine, Aarhus University, Aarhus, Denmark. Philip V Munch, Department of Clinical Epidemiology, Aarhus University Hospital, Aarhus, Denmark; Department of Clinical Medicine, Aarhus University, Aarhus, Denmark. Uffe Heide-Jørgensen, Department of Clinical Epidemiology, Aarhus University Hospital, Aarhus, Denmark; Department of Clinical Medicine, Aarhus University, Aarhus, Denmark. Lene H Iversen, Department of Clinical Epidemiology, Aarhus University Hospital, Aarhus, Denmark; Department of Clinical Medicine, Aarhus University, Aarhus, Denmark; Department of Gastrointestinal Surgery, Aalborg University Hospital, Aalborg, Denmark; Department of Clinical Medicine, Aalborg University, Aalborg, Denmark. Christian F Christiansen, Department of Clinical Epidemiology, Aarhus University Hospital, Aarhus, Denmark; Department of Clinical Medicine, Aarhus University, Aarhus, Denmark. Funding This study is supported by a grant from the Danish Cancer Society (grant no.: R361-A21253). Author contributions Mads Skovsgaard (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing—original draft, Writing—review & editing), Philip V. Munch (Conceptualization, Investigation, Methodology, Resources, Supervision, Writing—review & editing), Uffe Heide-Jørgensen (Conceptualization, Data curation, Investigation, Methodology, Supervision, Writing—review & editing), Lene H. Iversen (Conceptualization, Investigation, Methodology, Supervision, Writing—review & editing), and Christian F. Christiansen (Conceptualization, Funding acquisition, Investigation, Methodology, Resources, Supervision, Writing—review & editing) Disclosure The Department of Clinical Epidemiology of Aarhus University and Aarhus University Hospital receives funding for other studies from companies in the form of research grants to (and administered by) Aarhus University. The authors declare no other conflict of interest. Supplementary material Supplementary material is available at BJS Open online. Data availability The data underlying this study may be obtained from a third party. Researchers can apply for access to the data at www.dst.dk and https://www.rkkp-forskningsadgang.dk/. References Associated Data Supplementary Materials Data Availability Statement The data underlying this study may be obtained from a third party. Researchers can apply for access to the data at www.dst.dk and https://www.rkkp-forskningsadgang.dk/.