Abstract
Background Childhood, teenage and young adult (CTYA, 0-24 years) cancers are rare and diverse, making timely diagnosis challenging. Studies based on adult cancers suggest that the development and integration of clinical decision tools in primary care aid earlier cancer detection, yet, these have not been explored for CTYA cancers.
Aim To develop and validate a primary care-based risk prediction tool to identify CTYA who are at increased risk of cancer.
Methods and analysis Using the QResearch Database, a nationally representative primary care database, we will generate an open cohort of children, teenagers and young adults (0-24 years) who were registered with a GP between 1st January 1998 and 31st December 2019. CTYA will be followed up from the date at which the first cancer-relevant symptom is recorded in the records (index date) until the date of cancer diagnosis/6-months, whichever comes first. Candidate variables will include symptoms, signs, blood test results and demographic factors. Model derivation will include two approaches, Cox regression and logistic regression. Apparent performance of the derived model will be explored and subsequently internally-externally cross-validated to investigate performance heterogeneity and geographical transportability.
Introduction
Childhood and teenage cancer ranks as the 6th leading cause of total cancer burden worldwide and is associated with significant long-term morbidity1. Cancer is the commonest cause of mortality by disease among children and young people in the United Kingdom (UK)2. Delayed cancer detection is a known contributing factor3, with presentation at advanced stages recognised to reduce survival4. The UK has longer time-to-diagnosis across childhood cancers compared to other high-income countries5,6, as well as higher mortality rates across teenage and young adult cancers7. This is an ongoing concern for young people with cancer, who, in a recent national survey, highlighted early diagnosis research as one of their top ten research priorities8. This collectively highlights a clear health challenge and there is a pressing need to improve early detection of these cancers in the UK. Indeed, this is in line with the National Health Service (NHS) Long Term Plan (2019)9, UK Cancer Reform Strategy (2015)10 and Childhood Cancer and Leukaemia Group (CCLG) Strategic Plan (2020)11.
The non-specific presentation and relative rarity of childhood, adolescent and young adult cancers (CTYA) pose difficult diagnostic challenges to clinicians12 and increase the possibility of delays. This is particularly relevant to general practitioners (GPs) who encounter CTYA cancer patients at the earliest stages of the disease. National awareness initiatives, such as HEADSMART5, have been employed to address this challenge in the UK and although this initiative contributed to substantial improvements in diagnostic intervals in central nervous system (CNS) tumours, the national time-to-diagnosis target of 4 weeks has not been reached for all age groups13. Furthermore, GPs remain unconfident in diagnosing childhood cancers even after taking part in this initiative14. Similarly, recent findings of the “Accelerate, Coordinate, Evaluate” (ACE) programme demonstrated ongoing delays in referrals and cancer diagnosis in TYA15. Clearly, novel approaches need to be explored to supplement current pathways.
Computer-based clinical decision tools (CDTs) are increasingly being used in clinical settings, supporting medical decision-making where challenges such as diagnostic uncertainty are present16. Overall, CDTs have been reported to reduce diagnostic errors17, improve clinical practice and patient care17. In primary care settings, recent evidence suggests that technology-based CDTs provide the most successful interventions in reducing diagnostic inaccuracies18.The potential for CDTs in cancer diagnosis have been highlighted as an “area of extraordinary opportunity”, with promising developments seen in several adult cancers19. A recent systematic review20 has shown that these tools for cancer risk prediction have the potential to improve decision-making and clinical service outcomes, as well as one study showing reduction in time-to-diagnosis. Despite these advancements in adult cancers, however, CDTs have not been explored in CTYA cancers.
Accordingly, in this study, we plan to use QResearch Database, one of the largest GP electronic health record database in the UK, to explore ways to detect childhood and TYA cancers earlier by developing a novel GP-based risk prediction tool for CTYA cancers.
Methods and Analysis
Data sources and Study Population
Data Sources
QResearch Database is a nationally representative primary care database consisting of over 35 million anonymised health records from approximately 1300 general practices (GPs) in England (∼ 20% UK population)21,22. Records consist of patient-level demographic information (i.e., year-of-birth, sex, self-assigned ethnicity), as well as clinical information, including cancer diagnoses and clinical presentations. Primary care records are linked to hospital admission, civil registration and the National Cancer Registry data, where linkage is based on an individual patient’s anonymized NHS number. This number is valid and complete in 99.8% of primary care/civil registry data and 98% of hospital admissions data22.
Study Population
Model development and validation will use an open cohort of children, teenagers and young adults (from birth up to 25 years) who were registered with a GP within QResearch Database between 1st January 1998 and 31st December 2019.
Cohort entry will be the latest of date of registration with the practice plus 1 year, date on which the practice computer system was installed plus 1 year, and the study start date (1 January 1998) and for those who have cancer-relevant clinical features the first date in which the clinical feature was recorded. Exit from the cohort will be the earliest of 6 months following study entry date, 6 months following first recorded cancer-relevant symptom, or cancer diagnosis. Analyses will be restricted to CTYA who had a cancer diagnosis within 6 months or CTYA who had at least 6 months follow-up.
Cases will be defined as the commonest non-skin cancer diagnoses in this age group23 and will be categorised into subtypes according to the International Classification for Childhood Cancers (third edition, ICCC-3)24: 1) Leukaemias and myelodysplastic diseases, 2) lymphomas and reticuloendothelial neoplasms, 3) central nervous system and intraspinal tumours, 4) soft tissue and bone sarcomas, 5) abdominal tumours (renal tumours, neuroblastomas, hepatoblastomas) and 6) germ cell, trophoblastic and other gonadal tumours. Cases and their date of diagnosis will be identified through the National Cancer Registry. Cases with a diagnosis prior to study start date were excluded, as were those with the following pre-existing conditions linked to cancer: Down’s Syndrome, neurofibromatosis type I and II, ataxia telangiectasia, tuberous sclerosis, and Li Fraumeni 25-29. Incidence rates will be calculated for childhood (0-14 years) and TYA (15-24 years) and compared to available national incidence rates.
Outcome of Interest
The outcome of interest will be a diagnosis of the selected CTYA cancers within/at 6 months from presentation:
CTYA blood cancers (0-24 years): Any leukaemia/lymphoma diagnosis
CTYA solid cancers (0-24 years): Any CNS/sarcoma/abdominal tumour/gonadal germinal tumour diagnosis
Any CTYA who has a cancer diagnosis prior to cohort entry or after cohort exit will be excluded.
Selection of Clinical Features and Risk Factors (Candidate Predictor Variables)
Table 1 details all identified candidate predictor variables for model development. Cancer-associated clinical features in CTYA will be selected through previously published evidence available30-33. These clinical features (includes symptoms, signs and blood test results) will be identified through QResearch Database and results determined to be incorrect/outliers by the clinical team will be excluded. All blood test results will be categorised into normal and abnormal according to nationally available laboratory cut-off values34. Any clinical feature before cohort entry or after cohort exit will be excluded.
Record of sociodemographic risk factors on the study entry date will be used for data extraction. Ethnicity will be defined as self- or parent-reported ethnicity on primary care health records. Ethnic groups are recorded based on the 2011 Census of England and Wales in 2 broad categories (White, Other)35. Level of deprivation will be assessed through the Townsend deprivation score which is an area-level continuous score based on an individual’s postcode; factors that included unemployment, non–car ownership, non–home ownership, and household overcrowding, are measured for a given area of approximately 120 households, via the 2011 Census of England and Wales and combined to give a Townsend score for that area, with the first quintile representing the lowest deprivation level and the fifth quintile representing the highest deprivation level36.
Sample Size Calculations
Sample size calculations were carried out using ‘pmsampsize’ on Stata37. We set our time point at 6 months and used 15% of the maximum permitted Cox-Snell R-squared (derived from Riley et al., 202037) as there are no previous risk prediction models for its derivation. Cancer Research UK data were used to estimate incidence of cancers. We provide sample size calculations for the following cohorts: 1) blood cancers (leukaemias and lymphomas), 2) solid cancers (CNS, renal and hepatic tumours, soft tissue and bone sarcomas and neuroblastomas) (Table 2).
Previous studies using QResearch have identified a cohort of ∼5 million children and young adults38, with approximately 2 million children and 3 million teenagers and young adults. Altogether, this indicates that our study has sufficient sample size.
Model Derivation
We will explore two modelling approaches for model derivation:
Cox proportional hazards model
Logistic regression model
First, a full model will be fitted using all candidate predictors in all imputed datasets and combined using Rubin’s rules. Second, the pooled model will be used to select any categorical predictor with exponentiated coefficients >1.1 or <0.9 (at p<0.01) and any continuous predictors with significance of p<0.01. Third, these selected predictors will be used to refit the final model. This is to ensure both clinical and statistical magnitude of predictors are considered. Interactions between variables will be considered (based on a clinical plausibility) and interaction terms included within the model development process. Model coefficients will be combined, and in the case of Cox models with the baseline survival function, in order to calculate the linear predictor. Finally, model performance will be assessed through calculating the apparent discrimination, using Harrell’s C-index if cox regression and area-under-the-curve if logistic regression, and calibration, using calibration-in-the-large, calibration slope and smoothed calibration plots.
For cox regression models, proportional hazard assumptions will be assessed using Schoenfeld residuals.
Model Validation
The model derivation process will be validated using the internal-external cross validation approach39. This approach allows for the assessment of overall model performance as well as potential transportability of the model to a ‘distinct’ population and takes advantage of the availability of data from different geographical locations within QResearch Database (up to 10 locations across England). The summary of this approach is as follows: first, one geographical region will be ‘excluded’ whilst the model is developed using all other available regions. Second, data from the ‘excluded’ geographical region will be used to assess model performance using the aforementioned performance measures. These two steps will be repeated for each geographical region. Finally, random effects meta-analysis using the Hartung-Knapp-Sidik-Jonkmann method40 will be carried out to pool region-level performance measures and provide a pooled estimate of performance measures (i.e., discrimination and calibration).
Missing Data and other Statistical Consideration
We anticipate that there will be missing data for deprivation level (Townsend Quintile), ethnicity and blood test values and we consider these variables under the missing at random assumption41. For categorical variables we will use multinomial logistic regression, for ordinal variables ordinal logistic regression for imputation. 5 imputations will be carried out to strike a balance between % missingness and computational efficiency. Model coefficients will be pooled in accordance with Rubin’s rules. The imputation model will be inclusive of candidate variables and the outcome variable.
Any continuous candidate variable (e.g., age) will be assessed for nonlinearity and handled using fractional polynomials. Fractional polynomials will be fitted prior to imputation analyses. Clustering of participants within individual general practices will be accounted for using clustered standard errors.
Decision Curve Analysis
To assess potential clinical utility, a decision curve analysis will be used to compare standard net clinical benefit (i.e., the trade-off between the benefits of true positives and harms of false positives) using developed models with a scenario where no model is used.
Statistical Software
All analyses will be carried out using Stata (v17)42.
Patient and Public Involvement
National charities were approached to identify patient representatives and currently two representative young people have volunteered to provide input in study design, interpretation and dissemination of results.
Ethics and Dissemination
This project (OX94) has been approved by the QResearch scientific committee. The QResearch database annually obtains ethical approval from the East Midlands-Derby Research Ethics Committee (REC reference 18/EM/0400).
Data Availability
N/A
References
