Effect of surgery versus chemotherapy in pancreatic cancer patients: a target trial emulation

Abstract

Background

To estimate the causal effect of surgery vs chemotherapy on survival in patients with T1-3NxM0 pancreatic cancer in a rigorous framework addressing selection bias and immortal time bias.

Methods

We used population-based Danish health-care registries to conduct a cohort study emulating a hypothetical randomized trial to estimate the absolute difference in survival, comparing surgery with chemotherapy. We included pancreatic cancer patients diagnosed during 2008-2021. Exposure was surgery or chemotherapy initiated within a 16-week grace period after diagnosis. At the time of diagnosis, data of each patient were duplicated; one copy was assigned to the surgery protocol, and one copy to the chemotherapy protocol of the hypothetical trial. Copies were censored when the assigned treatment deviated from the observed treatment. To account for informative censoring, uncensored patients were weighted according to confounders. For comparison, we also applied a more conventional analysis using propensity score-based inverse probability weighting.

Results

We included 1744 patients with a median age of 68 years: 73.6% underwent surgery, and 18.6% had chemotherapy without surgery; 7.8% received no treatment. The 3-year survival was 39.7% (95% confidence interval [CI] = 36.7% to 42.6%) after surgery and 22.7% (95% CI = 17.7% to 28.4%) after chemotherapy, corresponding to an absolute difference of 17.0% (95% CI = 10.8% to 23.1%). In the conventional survival analysis, this difference was 23.0% (95% CI = 17.0% to 29.0%).

Conclusion

Surgery was superior to chemotherapy in achieving long-term survival for pancreatic cancer. The difference comparing surgery and chemotherapy was substantially smaller when using the clone-censor-weight approach than conventional survival analysis.

Pancreatic cancer has a dismal prognosis, with a 5-year survival below 10% (1,2). Observational studies generally report median survival of around 24 months for surgical treatment and 6-8 months for oncological treatment in patients with nonmetastatic disease (3-5). These survival estimates are somewhat lower than estimates from randomized trials (6,7), which, by comparison, generally include younger patients in better performance. Findings from randomized trials may therefore not be applicable at the population level.

Surgery is considered the preferred strategy for long-term survival, but few studies have compared the effect of surgery vs chemotherapy in patients with resectable or borderline pancreatic cancer. One study randomized 42 Japanese patients perioperatively to resection or radiochemotherapy, when the tumor was found to be resectable during laparotomy (8). Resection increased median survival by only 4 months compared with radiochemotherapy. A similar difference was found in a US-based propensity score-matched study of approximately 3800 patients (9). In a study of 81 Brazilian patients with resectable or borderline resectable tumors, surgery was associated with a median survival of 25 months compared with 9 months in the nonsurgery group (10).

Pancreatic cancer surgery is a high-risk procedure with in-hospital mortality of 1%-2% and morbidity of ∼30% (11). Because of the high risk of adverse outcomes of pancreatic cancer surgery, nonsurgical treatment may be preferred in some patients. Thus, knowledge on treatment effects of nonsurgical treatment is essential to inform patients of their different treatment options and expected outcomes. Observational studies of treatment effects are prone to immortal time bias, selection bias, and confounding (12-15). Thus, it is difficult to assess whether the superiority of surgery over chemotherapy is due to a true treatment effect or to study design-related biases. Observational data are increasingly used to generate evidence by emulating (hypothetical) target trials (16-18), which can avoid some of the common pitfalls that can lead to biased treatment effect estimates (19).

We used target trial emulation to estimate the effect of surgery compared with chemotherapy in patients with T1-3NxM0 pancreatic cancer to obtain clinically meaningful effect estimates of nonsurgical treatment.

Methods

Setting and data sources

We linked data from several nationwide health-care registries in Denmark to identify patients diagnosed with pancreatic adenocarcinoma during 2008-2021 (20-24). The registries are linked at the individual level using the Civil Personal Registration number, assigned to every Danish resident at birth or upon immigration (detailed descriptions in Supplementary Methods, available online).

Study design and population

The target trial

We first specified the target trial that would ideally be conducted to answer our study question, and then emulated this trial using observational data (25). Details of the components of the target trial and our emulation using observational data are specified in Supplementary Methods (available online). In brief, this would be a pragmatic, open-label randomized trial of patients with resectable or borderline resectable pancreatic cancer randomized to surgery or chemotherapy with the absolute difference in 3-year survival as primary estimand.

Emulation using observational data

Next, to emulate the target trial using observational data, we initially included all patients aged 18-85 years diagnosed with pancreatic adenocarcinoma in the Danish Cancer Registry or the Danish National Patient Registry during 2008-2021. The date of diagnosis served as the index date. We excluded individuals with less than 5 years of continuous residence in Denmark before diagnosis, those diagnosed at death/autopsy, and patients with unknown residence. We restricted the study population to patients meeting the inclusion criteria of the target trial (ie, patients eligible for both surgery and chemotherapy). Thus, we required that patients had a resectable tumor (resectability criterion) and were fit to undergo major surgery (operability criterion) and that no treatment was initiated before the date of diagnosis. We did not have information on tumor resectability, but we approximated this information by restricting to patients with T1-3M0 tumors (and any N-stage). To meet the operability criterion, we excluded patients with another incident cancer diagnosis (except breast cancer, nonmelanoma skin cancer, hematological malignancies, and duodenal/ampullary cancers as the latter would likely be misclassified pancreatic cancers) in the year before pancreatic cancer diagnosis and patients with a Nordic Multimorbidity Index (26) score greater than 15, albumin level lower than 25 g/L (as low albumin levels are associated with increased mortality in cancer patients) (27-29), or older than age 80 years (Supplementary Methods, available online).

Treatment protocols

In the target trial, treatment would be assigned randomly when eligibility was determined and would be delivered over a grace period after randomization. In our trial emulation, we used a 16-week grace period starting at the first date of pancreatic cancer diagnosis recorded in the Danish Cancer Registry or Danish National Patient Registry to define the two treatment protocols of surgery with no prior chemotherapy or chemotherapy without surgery (Supplementary Methods, available online). We restricted to chemotherapy codes recorded at oncological departments.

Covariates

Data sources and assessment windows of all covariates are shown in Supplementary Methods (available online).

Demographic data

We obtained information on marital status and municipality of residence, which are associated with receipt of treatment and survival (30,31). If information on marital status was unavailable (n = 117), patients were considered to be unmarried.

Tumor characteristics

Patients with M0/Mx status were excluded in case of histological verification of malignancy originating in the pancreas or upper gastrointestinal tract in cytology or biopsy (liver, lung, or peritoneum) within 30 days before the index date.

Comorbidity and lifestyle factors

We retrieved information on selected comorbidities for all patients, restricting to diagnoses recorded up to 5 years before the index date (Supplementary Methods, available online). To augment the assessment of comorbidities, we identified prescriptions used to treat the relevant comorbidities, restricting to 1-year lookback (Supplementary Methods, available online). We constructed a composite score of the overall comorbidity burden for each patient using the Nordic Multimorbidity Index, a validated comorbidity index designed to predict 5-year mortality in a Danish population (Supplementary Methods, available online) (26). We also obtained information on self-reported alcohol consumption and tobacco smoking within 180 days before the index date. We included blood samples drawn within 30 days before the index date. For patients with missing information on albumin levels, we used multiple imputation with predictive mean matching and 20 imputations to impute the value.

Statistical analyses

We analyzed the data using the clone-censor-weight approach (32). First, each individual was duplicated at the index date. One copy was assigned to the surgery protocol and the other to the chemotherapy protocol of the target trial. Second, the copies were artificially censored over time when they deviated from their protocol. Thus, copies assigned to the surgery protocol were censored at receipt of chemotherapy or if surgery was not performed within 16 weeks after diagnosis. Likewise, copies assigned to the chemotherapy protocol were censored if they received surgery or if chemotherapy was not initiated within 16 weeks after diagnosis (Supplementary Methods, available online). Follow-up ended at death, emigration, 5 years postdiagnosis, or November 10, 2023.

Because treatment was not randomized, potential confounding variables (age, sex, year of diagnosis, alcohol and smoking status, marital status, area of residence, and comorbidity; Supplementary Methods, available online) may have influenced the actual treatment decision, and thereby deviation from the assigned treatment protocol. Therefore, in our third step, we used inverse probability of censoring weighting to account for the selection bias introduced by informative censoring (25). Treatment-specific weights were calculated using a pooled logistic regression model including day of follow-up since the index date as a predictor, modeled as a restricted cubic spline with 3 knots. Weights were truncated at the 1st and 99th percentile (33). Throughout the grace period, alcohol, smoking, and comorbidity were updated daily as time-varying covariates. Missing information on smoking and alcohol was included as an indicator variable in the weighting models. To assess covariate balance, we calculated the standardized mean differences of potential confounders before and after weighting at the end of the grace period. A covariate with an absolute standardized mean difference between -0.1 and 0.1 was considered balanced (33). We used the Kaplan-Meier estimator to estimate and plot inverse probability of censoring weighted overall survival. Comparing the surgery protocol with the chemotherapy protocol, we reported differences in survival at 1, 3, and 5 years and median survival. All estimates are presented with standard error-derived 95% confidence intervals (CIs), calculated using bootstrapping with 200 repetitions. Statistical analyses were conducted using Stata 18.

Comparator analyses

To demonstrate the impact of different biases related to study design, we conducted 2 additional analyses on the noncloned cohort using survival analyses often used in observational studies. In the first analysis, we started follow-up at treatment initiation. We intentionally introduced selection bias in this analysis by excluding patients dying before treatment could be delivered. To mitigate confounding in this analysis, we used propensity score-based inverse probability weighting. Weights were calculated and truncated similarly to the target trial emulation. In the second analysis, we started follow-up on the date of diagnosis and used information on future treatments to define the surgery and chemotherapy groups at start of follow-up, intentionally introducing immortal time bias (12). In this analysis, we also excluded patients not receiving treatment, and no weighting or adjustments were performed to address confounding.

Sensitivity analyses

To examine robustness of our findings and to asses and quantify potential violations of the clone-censor-weight method, we conducted several sensitivity analyses as outlined in Supplementary Methods (available online).

Ethical considerations

Ethical approval was not required for this study, but it was registered at the Danish Data Protection Agency (j. no. 1-16-02-426-21) and complied with the General Data Protection Regulation.

Results

Descriptive characteristics

We identified 1744 patients meeting the study eligibility criteria. The median age was 68 years (interquartile range [IQR] = 62-73 years), and 52.1% were men (Table 1). Most patients (73.6%) underwent surgical treatment, 18.6% were treated with chemotherapy alone, and 7.8% had no treatment recorded. Nontreated patients were older and had higher levels of comorbidity compared with patients receiving treatment (Supplementary Table 1, available online). Of the 461 patients not receiving surgery during the grace period, 33 (7.2%) underwent surgical exploration or a presumed palliative procedure.

Table 1.

Descriptive characteristics of 1744 Danish patients diagnosed with T1-T3, M0 pancreatic adenocarcinoma during 2008-2021

	N (%)
Total	1744
Age, median (IQR)	68 (62-73)
Age group
<60 years	377 (21.6%)
61-70 years	668 (38.3%)
71-80 years	699 (40.1%)
Sex
Men	908 (52.1%)
Women	836 (47.9%)
Marital status
Married/registered partner	1150 (65.9%)
Unmarried	137 (7.9%)
Divorced	257 (14.7%)
Widowed	200 (11.5%)
Area of residence
Remote municipality	186 (10.7%)
Rural municipality	503 (28.8%)
Regional municipality	325 (18.6%)
Urban/metropolitan municipality	730 (41.9%)
Calendar period of diagnosis
2008-2010	302 (17.3%)
2011-2014	487 (27.9%)
2015-2017	538 (30.8%)
2018-2021	417 (23.9%)
Alcohol consumptionb
None	78 (4.5%)
1-14 units per week	158 (9.1%)
15-21 units per week	n ≤ 5a
>21 units per week	n > 10
Unknown	n > 1400
Alcohol consumption (composite)c
None/light	78 (4.5%)
Heavy	158 (9.1%)
Unknown	n > 1400
Tobacco smokingb
Non-smoker	182 (10.4%)
Current smoker	96 (5.5%)
Former smoker	26 (1.5%)
Unknown	1440 (82.6%)
Tobacco smoking (composite)c
No	171 (9.8%)
Yes	293 (16.8%)
Unknown	1280 (73.4%)
Nordic Multimorbidity Index, mean (SD)	2.7 (3.9)
Comorbidity
Stroke or other cerebrovascular	61 (3.5%)
Cardiac disease	317 (18.2%)
Hypertension	960 (55.0%)
Chronic lung disease	296 (17.0%)
Diabetes	438 (25.1%)
Chronic liver disease	16 (0.9%)
Kidney disease	21 (1.2%)
Alcohol-related disease	46 (2.6%)
Smoking-related disease	199 (11.4%)
Blood tests
Albumin, median (IQR)	35 (33-38)
Bilirubin, median (IQR)	29 (9-118)
CA19-9, median (IQR)	132 (36-453)
Tumor location
Head	1210 (69.4%)
Body	121 (6.9%)
Tail	90 (5.2%)
Other/multiple	52 (3.0%)
Unknown	271 (15.5%)
T-stage
T1	158 (9.1%)
T2	484 (27.8%)
T3	1102 (63.2%)
N-stage
N-	575 (33.0%)
N+	1031 (59.1%)
Nx	138 (7.9%)
AJCC stage
Ia	100 (5.7%)
Ib	174 (10.0%)
IIa	289 (16.6%)
IIb	858 (49.2%)
III	141 (8.1%)
Unknown	182 (10.4%)

a

Numbers collapsed for confidentiality. AJCC = American Joint Committee on Cancer; IQR = interquartile range.

b

Self-reported from the Danish Anesthesia Database.

c

Composite variables on the basis of self-reported use, prescription medications, and diagnosis codes (see Supplementary Methods, available online).

Target trial emulation (cloned cohort)

After weighting, most covariates were sufficiently balanced at the end of the grace period (Supplementary Figure 1, A, available online). The 3-year survival in the surgery protocol was 39.7% (95% CI = 36.7% to 42.6%) and 22.7% (95% CI = 17.1% to 28.4%) in the chemotherapy protocol, corresponding to an absolute difference in 3-year survival between the 2 protocols of 17.0% (95% CI = 10.8% to 23.1%). After 1 and 5 years of follow-up, the differences were 17.4% (95% CI = 11.3% to 23.5%) and 12.7% (95% CI = 7.3% to 18.1%), respectively (Table 2, Figure 1). Median survival was 25.9 (IQR = 11.8 to not reached) and 14.0 (IQR = 7.6-30.5) months in the surgery and chemotherapy protocol, respectively, corresponding to a difference of 11.9 (95% CI = 8.6 to 15.2) months.

Figure 1.

Survival curves comparing the surgery (solid line) and chemotherapy (dashed line) protocol in the target trial emulation using the cloned cohort. Curves are shown with corresponding 95% confidence intervals.

Table 2.

Survival estimates at 1, 3, and 5 years after diagnosis in 1744 pancreatic cancer patients (estimated using target trial emulation and the 2 comparator analyses)

	Survival, % (95% CI)
	1 year	3 years	5 years
Trial emulationa
Surgery	74.8 (72.4 to 77.3)	39.7 (36.7 to 42.6)	26.2 (23.3 to 29.2)
Chemotherapy	57.4 (51.5 to 63.3)	22.7 (17.1 to 28.4)	13.5 (8.9 to 18.2)
Difference	17.4 (11.3 to 23.5)	17.0 (10.8 to 23.1)	12.7 (7.3 to 18.1)
Comparator analysis (weighted)b
Surgery	76.6 (74.1 to 79.1)	41.1 (38.3 to 43.8)	27.3 (24.4 to 30.3)
Chemotherapy	50.2 (44.1 to 56.4)	18.0 (12.7 to 23.4)	12.5 (7.7 to 17.4)
Difference	26.4 (19.9 to 32.8)	23.0 (17.0 to 29.1)	14.8 (9.1 to 20.5)
Comparator analysis (nonweighted)c
Surgery	79.1 (76.1 to 81.5)	42.0 (39.1 to 44.9)	28.1 (25.2 to 31.0)
Chemotherapy	53.7 (48.1 to 59.3)	14.7 (10.7 to 18.7)	9.1 (5.9 to 12.4)
Difference	25.4 (19.3 to 31.5)	27.4 (22.4 to 32.4)	18.9 (14.6 to 23.3)

a

Target trial emulation using the cloned cohort with follow-up starting on the date of diagnosis. CI = confidence interval.

b

Propensity score-based inverse probability weighted analyses using the noncloned cohort with follow-up starting on the date of treatment initiation.

c

Non-weighted analyses using the noncloned cohort with follow-up starting on the date of diagnosis.

Comparator analyses (noncloned cohort)

In the propensity score-based inverse probability weighted survival analysis with follow-up starting on the date of treatment initiation, the 3-year absolute difference in survival was 23.0% (95% CI = 17.0% to 29.1%), comparing surgery with chemotherapy (Table 2). Median survival was 27.1 (IQR = 12.5 to not reached) months for surgery and 11.9 (IQR = 6.7-23.9) months for chemotherapy (Figure 2). Supplementary Figure 1, B (available online) shows covariate balance in the noncloned cohort before and after weighting. In the analysis intentionally introducing immortal time bias, the 3-year absolute difference in survival was 27.4% (95% CI = 22.4% to 32.4%; Table 2). In this analysis, median survival was 28.4 (IQR = 13.7 to not reached) months for surgery and 12.9 (IQR = 7.6-23.2) months for chemotherapy (Figure 3).

Figure 2.

Survival curves comparing the survival comparing surgery (solid line) and chemotherapy (dashed line) in the comparator analysis (propensity score-based inverse probability weighted analyses using the noncloned cohort with follow-up starting on the date of treatment initiation). Curves are shown corresponding with 95% confidence intervals.

Figure 3.

Survival curves comparing the survival comparing surgery (solid line) and chemotherapy (dashed line) in the comparator analysis (nonweighted analyses using the noncloned cohort with follow-up starting on the date of diagnosis). Curves are shown with corresponding 95% confidence intervals. The “chemotherapy” group in this analysis is particularly prone to immortal time bias as evident in the horizontal line in the first period after start of follow-up.

Sensitivity analyses

The difference in 3-year survival attenuated slightly when using more restrictive operability criteria (11.8%) and when including tumor stage in the model (12.1%). There was little impact when using a less (15.4%) restrictive approach to operability, excluding patients with prior pancreatic surgery (18.7%), censoring at palliative surgery/metastasis diagnosis, not censoring at chemotherapy receipt in the surgery protocol (16.6%), excluding patients with missing albumin levels (19.2%), or using different grace periods (17.7% to 18.8%). The difference increased when restricting to T1-T2 cancers (30.5%) or restricting to patients diagnosed in 2018-2021 (25.5%); full details in Supplementary Table 2 (available online).

Discussion

We applied the target trial emulation framework using observational data from Danish registries to estimate the effect of surgical treatment compared with chemotherapy on survival in T1-3NxM0 pancreatic cancer patients. We observed a 3-year survival of 39.7% in operated patients and 22.7% in patients treated with chemotherapy, corresponding to an absolute difference of 17.0%. After 5 years, the difference decreased to 12.7%.

Surgery is considered the only chance for long-term survival in pancreatic cancer, but it is a high-risk procedure. Evidence of the effect of nonsurgical treatments is therefore important to inform patients. A randomized trial by Imamura et al. (8) from 2004 comparing surgery with radiochemotherapy in pancreatic cancer patients found a 4-month difference in median survival. All patients in the radiochemotherapy arm received oncological treatment as planned. This is surprising, as up to half of the patients undergoing surgical exploration never initiate chemotherapy (34, 35). The study was conducted at a time with limited oncological treatment options, and the difference may be smaller with multiagent chemotherapy available. This was examined by Landa et al. in a propensity score-matched cohort study of 5146 patients (9), where the difference in median survival was similar to the study by Imamura et al. (4-5 months in both studies). The Landa et al. study observed a median survival of 14 months in operated patients, which is lower than expected (36) and likely attributed to exclusion of patients receiving neoadjuvant/adjuvant oncological treatment. Whereas adjuvant chemotherapy improves the prognosis considerably (37), the use of neoadjuvant chemotherapy in patients with resectable pancreatic cancer is controversial (38, 39). Our observed difference in median survival of 11.9 months is higher than in the studies by Imamura et al. and Landa et al. but comparable with the 13 months observed in a study of 81 stage I-II pancreatic cancer patients, of which 14 received chemotherapy only (10).

We used the target trial emulation framework to estimate the effect of surgery and chemotherapy for T1-3NxM0 pancreatic cancer. We found that the difference between surgery and chemotherapy was smaller than using methods introducing immortal time or selection bias. This finding indicates that commonly occurring methodological pitfalls in observational research can bias survival estimates. To obtain causal estimates, three main assumptions must hold: exchangeability (no unmeasured confounding and no informative censoring), positivity (a nonzero probability of being assigned to any of the treatment arms), and consistency (well-defined exposures) (40).

The possibility for unmeasured confounding should be considered, although we had sufficient balance for most covariates. First, we did not have information on performance status, which is associated with both prognosis and treatment decision (41). To address this, we included information on albumin as a proxy for performance status, as low albumin levels are associated with increased mortality in cancer patients (27-29). However, albumin has not been validated as a proxy, and residual confounding is a possibility. There was little effect on the 3-year survival when excluding patients with imputed albumin levels (19.2% vs 17.0% in the main analysis). Second, we had limited information on socioeconomic position, which may be associated with curative-intent surgery and improved survival (30,31), at least in nonuniversal health-care systems (42). To address this concern, we included area of residence and marital status in the weighted models as both are strongly associated with both socioeconomic position and receipt of treatment. Third, to mitigate underreporting of lifestyle-related comorbidity (43), we used information on prescription drug use and self-reported exposure to alcohol and smoking to augment comorbidity registration. Informative censoring was mitigated using inverse probability of censoring weighting. Where possible, we included continuous variables as restricted cubic splines (44).

Almost three-quarters of the study population received surgery during the 16-week grace period. The reason for not offering surgery may be nonresectability or nonoperability, both of which would contribute to violating the positivity assumption. As we did not have information on resectability or operability available, we applied strict criteria to approximate these assumptions. However, the difficulty in assessing tumor resectability in pancreatic cancer can introduce concerns about nonpositivity (45,46). To examine this concern, we conducted a sensitivity analysis, where we censored patients if they experienced metastases or had palliative/explorative surgery without resection before treatment could be initiated. Estimates from these analyses did not differ substantially from the main analyses. Only 7.2% of patients not undergoing curative-intent treatment experienced this, suggesting that tumor progression/nonresectability detected intraoperatively was an explanation for not offering surgery in a minority of the patients. Restricting to T1-T2 tumors resulted in an increase in the difference between surgery and chemotherapy. This increase may be because surgery is more effective than chemotherapy in low-stage tumors but also may be because of different baseline risks in the two populations (Supplementary Methods, available online for more details). Tumor stage was not included in the main analysis, as information on N-status was unknown in one-quarter of patients not receiving surgery, likely because of the low sensitivity of CT scans for detecting lymph node metastases (47,48). In the analysis including information on tumor stage, the difference in 3-year survival attenuated slightly. However, estimates from these sensitivity analyses should be interpreted with caution because of the poor reporting of lymph node metastases in nonoperated patients and the low number of patients with T1-T2 tumors. In our sensitivity analyses, applying less restrictive operability criteria, the estimates were comparable with that of the main analysis. When we used more restrictive operability criteria, the estimate attenuated slightly, but the precision was low. Surgery is generally underutilized even in operable patients with resectable tumors (49,50).

Although there has been little development in the surgical procedures during the study period, different chemotherapy regimens have been used. Oncological treatment of pancreatic cancer in Denmark generally conforms to national guidelines, although some variations have been shown (3), which could affect the consistency assumption.

Our study provides important estimates of the effect of surgery vs chemotherapy for patients with T1-3NxM0 pancreatic adenocarcinoma. Our findings support that some of the beneficial effect of surgery over chemotherapy in these patients can be attributed to selection bias and immortal time bias. Our findings are useful for clinicians to inform patients about nonsurgical treatment options.

Data availability

The data underlying this article cannot be shared without violation of Danish law. Researchers can obtain access to the data through a formal application on the Danish Health Data Authority website (www.sundhedsdatastyrelsen.dk).

Author contributions

Jakob Kirkegård, MD (Conceptualization; Data curation; Formal analysis; Funding acquisition; Writing—original draft; Writing—review & editing), Charles Gaber, PhD (Conceptualization; Writing—original draft; Writing—review & editing), Uffe Heide-Jørgensen, PhD (Conceptualization; Supervision; Writing—original draft; Writing—review & editing), Claus Wilki Fristrup, MD, PhD (Conceptualization; Writing—original draft; Writing—review & editing), Jennifer Lund, PhD (Conceptualization; Methodology; Supervision; Writing—original draft; Writing—review & editing), Deirdre Cronin-Fenton, PhD (Conceptualization; Supervision; Writing—original draft; Writing—review & editing), Frank Viborg Mortensen, MD, DMSc (Conceptualization; Funding acquisition; Resources; Writing—original draft; Writing—review & editing).

Funding

This work was supported by Spogards Fond/the Danish Cancer Society (grant number R302-A17201-B2567); Grosserer M. Brogaard og Hustrus Fond (grant number 5523-TL-ne); Aage og Johanne Louis-Hansens Fond (grant number 18-2B-3762); and Lundbeck Foundation (grant number R403-2022-1251).

Conflicts of interest

The authors declare no conflicts of interest.

Acknowledgments

The funding bodies had no role in the design and execution of the study or the decision to publish.