Semi-automated monitoring of deviations from historical incidence trends for enhanced timeliness in cancer control

Abstract

Background

Population-based cancer registries aim for timely reporting of cancer incidence and rapid identification of unexpected patterns to support cancer control. To facilitate large-scale, systematic, first evaluation of incidence data, a semi-automated tool was built that flags statistically significant deviations from historical trends, which may merit further investigation.

Materials and methods

The tool was developed in the open-source programming language R and uses a CSV file to specify input parameters. For each row, observed or age-standardised incidence rates are calculated, and corresponding graphical outputs, with various stratifications and substratifications, are produced. A flexible regression model is fitted to all but the most recent incidence year, with an automated algorithm determining optimal knotpoint placement. The model is then extrapolated to estimate the expected incidence for the final year, which is statistically compared with the observed value. The model is visually represented, including confidence bands, in the graphical output. Results are exported as PNG files organised in hierarchical folders, along with a structured Quarto HTML report containing all figures.

Results

The modelling approach captured complex temporal patterns without overfitting and reliably identified deviations from historical trends. This high-throughput tool is highly customisable and can be extended to analyse stage-specific incidence trends or stratifications by other categorical variables.

Conclusions

This semi-automated tool enables efficient first-line visual evaluation of newly available registry data to highlight emerging trends potentially warranting further investigation from either a data quality or public health perspective. The code is available from the authors upon request. Output is highly customisable via CSV input and can be stratified on multiple levels.

Graphical abstract

Highlights

• •High-throughput, semi-automated analysis of cancer incidence, ASR.
• •Flexible model fit to historical data.
• •Model extrapolated to compare expected and observed values for the last incidence year.

Introduction

Population-based cancer registries (PBCRs) maintain extensive datasets with regular updates. They have the social and often legal responsibility to report on this data as soon as possible. Considerable attention has been devoted to improving ‘timeliness’ within PBCRs, particularly by reducing the interval between diagnosis and the availability of cancer incidence reporting.1, 2, 3, 4 Enhancing the accessibility and availability of data is also a key objective of most PBCRs.⁵ However, for advances in timeliness and accessibility to be of value in cancer control, the subsequent analysis and interpretation of these data must also be timely.

A central contribution of PBCRs to cancer control is the proactive monitoring of emerging trends and the rapid identification of unexpected deviations from historical patterns, which may signal public health concerns.⁶ This role became especially clear during the coronavirus disease 2019 (COVID-19) pandemic, when lockdown measures reduced access to diagnostic services and PBCRs were called upon to quickly quantify the impact on cancer detection. Registries were also uniquely positioned to assess, at the population level, whether diagnostic delays had resulted in stage shifts. Timely trend detection is equally important when evaluating public health interventions such as newly implemented screening programs, where identifying shifts towards earlier-stage disease helps determine program effectiveness.^7,8

Not all stage shifts (changes in the distribution of disease at diagnosis) or stage migrations (changes in the proportion of tumour, node, or metastasis categories after the introduction of new means of assessing disease, such as those caused by advances in diagnostic technology) can be anticipated. The introduction of new diagnostic tools or subtle updates to staging definitions may lead to shifts in stage or basis of diagnosis, possibly in a limited patient population, that remain unnoticed without detailed longitudinal assessment. Similarly, unexpected changes in the distribution of cancer incidence, such as shifts in age-specific burden, may not be detected without comprehensive evaluations of incidence by age group and cancer (sub)type over time. Routine monitoring of incidence trends, stratified by stage, age, basis of diagnosis, region, and other key variables, also supports the identification of registration irregularities and enables ongoing evaluation of the completeness of these variables. In this context, such monitoring may equally serve as a quality control assessment.

Assessing deviations from historical trends is inherently challenging. During the COVID-19 pandemic, several approaches were attempted. The simplest involved comparing incidence in 2020 directly with 2019, or with the average from 2017 to 2019, but such comparisons fail to account for expected year-to-year changes driven by population ageing or migration. Age-standardised rates address demographic shifts but still overlook other anticipated temporal changes in incidence, such as those arising from evolving exposure to risk factors. To overcome some of these limitations, we previously applied a method that extrapolated trends from 2017 to 2019 to generate expected values for 2020.9, 10, 11 While pragmatic at the time, this approach is highly sensitive to one-off fluctuations in the data, making it difficult to apply reliably in the years immediately after the pandemic. When using a longer reference period, the application of a linear model to derive an expected value becomes problematic, as cancer incidence data often contain substantial temporal fluctuations.

Our aim was to develop a general-purpose, semi-automated monitoring tool that could provide rapid visualisations with automatic identification of (significant) deviations from historical trends for processing large amounts of diverse PBCR data. These initial detections would serve as hypothesis-generating signals, enabling proactive follow-up to determine whether emerging patterns reflect changes in the underlying cancer burden, or rather in clinical practise or registration procedures. To meet these objectives, the tool needed to clearly present temporal data and flag statistically significant changes from expected trends. The model underlying the historical trend had to incorporate all available past data while remaining flexible enough to accommodate natural undulations in cancer incidence without overfitting. Finally, we sought to develop the tool using open-source software so that it could be adopted by any PBCR, including those with limited resources.

Materials and Methods

Belgian Cancer Registry data

The Belgian Cancer Registry (BCR) covers all of Belgium and is estimated to be 98% complete from 2004 onwards.¹² Belgium comprises three main regions: Flanders, Wallonia, and the Brussels-Capital Region and it has a population of ∼12 million inhabitants. Reporting of cancer cases to BCR is mandatory for all oncological care programs and laboratories for pathological anatomy in Belgium based on the Coordinated law of 10 May 2015 (art. 138). In 2023, there were 77 332 new diagnoses of invasive cancers (excluding non-melanoma skin cancers) registered at BCR. Clinical and pathological staging was recorded according to the International Union Against Cancer (Union Internationale Contre le Cancer) TNM (tumour–node–metastasis) Sixth Edition (2004-2009)¹³, Seventh Edition (2010-2016)¹⁴ and Eighth Edition (2017-2025)¹⁵. Tumour characteristics were registered using International Classification of Diseases for Oncology, Third Edition (ICD-O-3) codes and converted to International Classification of Diseases, Tenth Revision (ICD-10) codes for classification and reporting.^16,17 Estimates for the n+1 incidence year at BCR were based on pathology notifications only.¹⁸

All analyses were carried out using RStudio running R 4.4.1 (The R Foundation for Statistical Computing, Vienna, Austria).

Analysis of incidence counts by the monitor tool

Annual incidence counts (crude numbers, N) are plotted with year on the x-axis and N on the y-axis using ggplot 2.0 in R. Stratification can be carried out by sex, age group, or region. Total incidence and stage (or another categorical variable)-specific incidence are plotted separately. Error bars [95% confidence intervals (CIs)] are added to each incidence value using the method of Daly.¹⁹

A Poisson model with a restricted cubic spline is fitted to the incidence data, excluding the most recent incidence year and any manually excluded years (e.g. 2020 could be excluded if the drop in diagnoses during the COVID-19 pandemic strongly impacted the prediction for 2021). The model takes the incidence count as the dependent variable and the incidence year as the independent variable, and the identity link function is applied using the R function glm. The natural spline for the incidence year is created with the function ns from the R package splines. The 95% CIs on the regression line are obtained from the predict.glm function. If a negative prediction is obtained, a new Poisson model with a log-link is applied to prevent negative predictions. For stage (or other variable)-specific incidence, the curve for each level is fitted independently.

The number and position of the spline knotpoints are automatically added to the model. To prevent overfitting, knotpoints are required to be at least 3 years apart and cannot be placed in the first or last 3 incidence years included in the calculation of the model. A looping function tests all possible knotpoints and combinations and selects the combination of knotpoints resulting in the lowest Akaike information criterion (AIC). If an additional knotpoint lowers the AIC by <2 points, the model with fewer knotpoints is selected. The curve is plotted including confidence bands, together with the observed incidence counts.

To create a predicted value for the most recent incidence year, the curve is extrapolated by 1 year. The observed and predicted incidence counts for the most recent year are compared using a two-sample z-test, as both values are independent of each other. The square root of the observed count is taken as the standard error for the observed count, while the standard error for the predicted count follows from the model prediction. A p value under the hypothesis of no difference is reported on the figure.

Age-standardised incidence rate

The age-standardised rate (ASR) and its 95% CI are obtained using the European Standard Population (ESP2013) or the World Standard Population reference population. Analysis of the ASR trends is carried out as for the incidence count trends, with minor modifications. A weighted linear regression model with the ASR as the dependent variable, assuming a normal error, is used, with the incidence year as the independent variable and an identity link. The weights are equal to the inverse of the variance on the ASR. If a negative prediction is obtained, a new weighted linear regression model with a log-link is applied to prevent negative predictions.

Automation

The desired output is highly customisable. Within a CSV input file, desired parameters including cancer type selection on ICD-10 code or ICD-O-3 topography and morphology codes, detailed subselections, stratifications, incidence years to include or exclude, desired extra year labels via the keyyears variable (e.g. year of start screening), axis labels, desired colour schemes, etc. are specified, with one line per desired set of analyses (for definitions of all variables in the input file, see Supplementary Table S1, available at https://doi.org/10.1016/j.esmorw.2026.100746). The input file allows for additional specific filtering using a Tidyverse ‘pipe’ [%>% filter(optional code)] such that selections can be fully customised using the initial_filters variable.

The input CSV file is imported into the monitor tool. The code loops over the data preparation and creates several graphical outputs for each row in the CSV file, beginning with (optionally) total incidence/ASR and incidence/ASR per level of the general stratification variable (generalvar, can be sex, region, or age group—must be likewise defined in the population file for calculating ASR). Subsequently, graphics are produced for each level of the generalvar, showing the incidence/ASR per level of the selected additional stratification variable (stratvar, can be any selected categorical variable in the database, with a maximum of seven levels, e.g. stage, basis of diagnosis, etc.). The p-values of the comparison between the observed and predicted values for the last incidence year are displayed on the graphs. An example call for female breast cancer with generalvar as age group and stratvar as clinical stage can be found in Supplementary Table S2 and File S1, available at https://doi.org/10.1016/j.esmorw.2026.100746.

The tool generates an HTML Quarto report that compiles all produced graphics and includes a table of contents enumerating each figure. In addition, it exports individual PNG files to the designated output directory, organised into date-stamped folders, with one folder corresponding to each line in the input CSV file.

Results

The structure of the output graphics remains generally consistent (Figure 1).

• 1)The observed incidence counts (N)/ASR values are plotted with their CIs.
• 2)The model is plotted based on data for n − 1 incidence years, including confidence bands.
• 3)The model is extrapolated to the last incidence year.
• 4)The results of the z-test comparing the observed and predicted values of the last incidence year are reported.
• 5)Any additional annotations are added.

Figure 1.

Example output from the monitoring tool for the total number of lung cancer diagnoses in Belgium from 2004 to 2023. (1) Observed incidence values with confidence intervals. (2) Plotted flexible model based on data for n − 1 incidence years, including confidence bands. (3) Extrapolation of the model to the last incidence year to generate the expected value. (4) The results of the z-test comparing the observed and predicted values of the last incidence year are reported. (5) Additional customisable annotations.

The tool was able to fit complex trends while avoiding overfitting; see the example of lung cancer (C34; Figure 2) and prostate cancer (C61; Figure 3). For lung cancer, the tool detected a significant decline in the incidence of stage IV lung cancer in females for the year 2023 compared with the expected number based on historical (2004-2022) trends (Figure 2A). No significant difference was observed for microscopically confirmed diagnoses (Figure 2B). Furthermore, the tool recovered readily from abrupt stage migrations (Figure 3B), as demonstrated by the change in 2010 from TNM6 to TNM7 for prostate cancer, with a profound migration from stage II to stage I prostate cancer. The COVID-19 decline had little to no impact on the overall fit of the curves. As such, the year 2020 was ultimately not excluded from the models.

Figure 2.

Example output from the monitoring tool for the number of lung cancer diagnoses in females by (A) clinical stage and (B) basis of diagnosis in Belgium from 2004 to 2023. A significant decline in stage IV lung cancer is observed in 2023, together with a significant deviation from the descending trend for unknown stage (X). The basis of diagnosis does not deviate significantly from historical trends. Points represent the number of diagnoses per year, per stage, with confidence intervals. Curves represent the individual flexible models fitted to the data from 2004 to 2022 and extrapolated to 2023, including confidence bands.

Figure 3.

Example output from the monitoring tool age-standardised incidence rates (ESR2013) for prostate cancer in Belgium from 2004 to 2023. (A) The model accurately follows the declining and subsequently increasing trends, without overfitting. A significant decline relative to the trend is noted for 2023 and 2020. (B) With the implementation of Union for International Cancer Control (UICC) TNM 7¹⁴ staging, a stage migration is observed from stage II to stage I. Points represent the age-standardised incidence per year (per stage in B) with confidence intervals. Curves represent the individual flexible models fitted to the data from 2004 to 2022 and extrapolated to 2023, including confidence bands. Values for 2024 are estimates based on pathology notifications only.

For colorectal cancer, roll-out of organised screening began in the Flemish part of Belgium in 2013 (complete roll-out in 2014; targeted age range gradually expanded from 56-74 in 2014 to 50-74 from 2020 onwards). As an example, for rectal cancer, a spike in invasive tumours was observed specifically in the screening age group in 2014, followed by a slowly declining incidence in both the screening and older age groups (Figure 4A). Given this single spike, the year 2014 was omitted from the modelling of the trends. The stage distribution, specifically for the screening age group, demonstrated a peak in stage I tumours in 2014, followed by a general increase in stage I tumours and declining trends in stage III and IV tumours, and a sharp decline in tumours of unknown stage (X) (Figure 4B).

Figure 4.

Example output from the monitoring tool for age-standardised incidence rates (ESR2013) for rectal cancer by age group for both sexes (A) and clinical stage for both sexes in the screening age group 50 to 74 years old (B) in Belgium from 2004 to 2023. (A) The implementation of organised screening in the Flemish region [‘Start screening (Flanders)’] is observed with a peak in cancer incidence in the screening age group (50-74 years). (B) An increasing trend in stage I is observed starting in approximately 2010, followed by a declining trend in stages III and IV in the screening age group (50-74 years). Points represent the age-standardised incidence per year (per stage in B) with confidence intervals. Curves represent the individual flexible models fitted to the data from 2004 to 2022 and extrapolated to 2023, including confidence bands.

The tool is highly customisable, and analyses can be made for a broad range of stratifications and general groupings. For example, different treatment schemes for non-small-cell lung cancer were analysed based on the general stratification of patients above and below 70 years of age (Figure 5). Treatment patterns remained relatively stable in the younger age group (Figure 5B), although a switch in the preference for surgery with additional systemic and/or radiotherapy to surgery only was seen in recent years. In the older age group (Figure 5C), treatment patterns were more variable over time, with a decline in the number of patients receiving no treatment.

Figure 5.

Example output from the monitoring tool for the treatment of non-small-cell lung cancer (NSCLC) in 2004-2022. (A) The incidence counts of NSCLC are increasing in people aged ≥70 years. (B) Patients aged <70 years are typically treated with a combination of systemic therapy and radiotherapy (RT); few patients receive no curative treatment or RT only. (C) For patients aged ≥70 years, no active anticancer treatment is more common than surgery alone or RT alone, or surgery combined with either RT and/or systemic therapy. Points represent the number of diagnoses per year (per treatment in B and C) with confidence intervals. Curves represent the individual flexible models fitted on the data from 2004 to 2021 and extrapolated to 2022, including confidence bands.

Discussion

Tool performance

We developed a semi-automated tool for rapid initial evaluation of cancer incidence, for both incidence counts and age-standardised incidence rates, overall or for specific subgroups, including stratified analyses such as stage-specific analyses. Compared with creating individual figures and analyses based on expected deviations, semi-automated, high-throughput initial evaluation by cancer type, age group, and other possible stratifications enables unbiased review of the available data to identify issues requiring further follow-up. This tool was successfully implemented at BCR; a first version of this tool was used to provide figures for analysis just 1 day after the closure of our databank for the incidence year 2023.

The model used in our tool smooths trends in cancer registry data without losing important fluctuations. Compared with the average annual percent change, which is fitted on the log scale and based on piecewise linear models,²⁰ our model uses a spline, which is more flexible and smooths general trends in the data over time, making long-term interpretation possible. The confidence bands provided on the output graphics also allow for the detection of historical deviations and give a visual cue of the variability of the data. At moments of sharp change, such as the implementation of organised colorectal cancer screening in the Flemish region (Figure 4), it may be necessary to omit an incidence year from the model to optimise the fit.

The flexible model with automatic assignment of knotpoints performed well for the purpose of rapid evaluation of potential deviations; no manual adjustment of knotpoints was required to fit our data. The model was able to adequately track the complex trends observed in PBCR data. What join-point analysis offers, which is not available in our model, is an average trend value and an exact indication of when trends in data change. Knotpoints in our model identify places where the parabolic trend has changed, but their interpretation is complex. Nevertheless, given that this tool is intended to identify deviations that may merit further investigation, the expectation is that an observed change in trend would be followed by further complementary analyses, such as join-point.

Interpretation of results

We wish to emphasise that the output of this tool is meant to assist in the rapid initial evaluation of cancer registry incidence data and is intended to provide an 'initial signal’ for further investigation and interpretation. The output of graphics displaying the temporal trends was an intentional decision to invite the end user to visually interpret the fit of the model and not only rely on the statistical testing.

Interpretation of stage shift/migration needs to be undertaken with caution and must account for trends in the unknown stage. The unknown stage can arise from either a patient being unfit/refusing staging interventions, in which case, the stage is then truly unknown, or complete staging may have been carried out, but the information was not transmitted to the cancer registry. Collection of complete staging data has been identified as a common challenge for PBCRs.⁵ Availability of staging information may be augmented by interpreting unknown cN or cM values as cN = 0 and cM = 0, respectively.²¹ In Belgium, we have observed a large improvement in the quality of staging information in the cancer registry database in the early 2010s. The decline in the unknown stage, however, made interpretation of changes in stage distribution difficult. For lung cancer (Figure 2), part of the rise in stage IV until 2022 may be because of improved data completeness and may be the complement to the decline in stage X. As a plateau took place in the unknown stage, stage IV may be likewise expected to plateau, and the observed significant decline in stage IV in 2023 may be the beginning of this plateau.

When interpreting the stage, it is crucial to distinguish the clinical (c) stage from the pathological (p) stage, and from the posttreatment pathological (yp) stage. It is common practice to use ‘one stage’ (combining available clinical and pathological staging information) in cancer reporting to augment data availability; however, in the context of this tool, we would strongly discourage this practice. It is particularly important, given changing trends in the use of neoadjuvant therapy and surgical approach (compared with other treatments with curative intent), to distinguish between non-operated patients, primary surgery, and surgery after neoadjuvant treatment when interpreting (y)p/c stage differences.

Future improvements

In future iterations of this tool, we intend to incorporate the ability to follow temporal trends in observed survival. Currently, only 60% of cancer registries publish survival reports.⁵ It is possible to assess mortality data using the currently available models for ASR; however, survival data, which allow for cancer type-specific analyses, will undoubtedly be of additional benefit. For example, although little difference in stage at incidence was observed after the COVID-19 pandemic in Belgium,9, 10, 11^,22,23 it is possible that subtle within-stage shifts or moderate adaptations to treatment plans could lead to higher risks of recurrence and death. By extensive and detailed monitoring of stage-specific survival by cancer type, such issues could be detected in a timely manner and acted upon.

Conclusion

Expanding the use of new tools and reducing time to reporting have been identified as goals by the majority of PBCRs.⁵ We developed a high-throughput monitoring tool that is intended for rapid initial evaluation of incidence data to improve the timeliness of evaluation of PBCR data. The code is available from the authors for any PBCRs wishing to apply it in their own setting for quality monitoring and cancer control.

Ethical Approval

This observational study does not require approval from an Ethical Committee in Belgium. The study was carried out within the legal and ethical framework of the Belgian Cancer Registry (BCR). Based on the Coordinated law of 10 May 2015 (art. 138) (Het Belgisch Staatsblad. [fgov.be]; https://www.ejustice.just.fgov.be/doc/rech_n.htm), the BCR has a legal task to collect data on cancer, subject it to quality control, process and analyse it, encrypt and store it, and report on it. Its activities are monitored and evaluated by the Advisory Committee of Users, an independent committee whose duly qualified members are appointed by law.

Disclosure

The authors have declared no conflicts of interest.