Design and setting
A retrospective cohort study in 21 primary-care practices in South-West England. Data were extracted for the RAPCI Study (Rapid COVID-19 intelligence to improve primary-care response), a mixed-methods study on the rapid change to remote consulting in the initial months of the pandemic; primary results are reported elsewhere.4
Data
Routinely collected and anonymised data were provided by One Care, the GP federation in Bristol, North Somerset, and South Gloucestershire. All practices use the EMIS electronic medical records system. Data included demographics (age, sex, ethnicity, and deprivation), clinical characteristics (mental health and shielding status), and all consultations and clinical codes associated with consultations, added to the system by clinical staff between February 2019 and July 2020 inclusive. All patients registered in July 2020 were included. For the analyses, April-July 2020 (i.e. the period following UK lockdown) was compared to April-July 2019.
Consultations
Consultations were defined as an interaction between a patient and a GP, nurse, or paramedic working in general practice. Consultations recorded by administrators or other health care professionals, and any administrative tasks, were excluded. Remote consultations were those completed by telephone, video, or e-consultation; face-to-face consultations were in GP practices or visits to patients’ homes (see appendix 1).
Outcomes
Pre-existing code lists10,11 were used to identify potential cancer indicators associated with a consultation. Indicators were collated from clinical features of undiagnosed cancer (symptoms, signs, abnormal test results or diagnoses) listed in the National Institute for Health and Care Excellence guidance on the recognition and referral of suspected cancer (NG12)12, using robust methods11 (see appendix 1).
Individual potential cancer indicator were categorised by the percentage of patients reporting them in April-July 2019: most commonly (≥0.5%), less commonly (0.1% to <0.5%), rarely (0.02% to <0.1%), and very rarely reported (<0.02%). We separate these because the most commonly reported indicators include symptoms which often indicate minor illness rather than cancer (e.g. cough), whereas the less commonly/rarely reported indicators were more likely to be associated with cancer (e.g. weight loss, lumps and masses).
Explanatory variables
Age (in July 2020) was categorised: 0-4, 5-17, 18-49, 50-69, 70-84, and 85+ years old. Deprivation quintiles were calculated using index of multiple deprivation score (IMD) deciles recorded in patient records, based on Lower Super Output Areas of residence. Ethnicity was derived by mapping descriptions from primary-care records to five categories: white, Asian, black, mixed, and other (see appendix 1). Presence of a mental health condition included severe mental illness (defined according to the Quality and Outcomes Framework rules13), diagnosed depression, or prescribed anti-depressants (excluding tricyclics) in the three months prior to July 2020. Sex and shielding status (as of July 2020) were obtained directly from primary-care records.
Statistical analysis
Number and percentages of patients reporting potential cancer indicators are presented. Further, consultation rates (and percentages with indicators) are reported per 1,000 registered patients. Practice list sizes were based on July 2020 data, and adjusted to account for historic list sizes using NHS digital data14 (‘adjusted list size’; see appendix 1).
Changes in proportions of patients presenting (per practice) with any potential cancer indicator in April-July 2020 compared to April-July 2019 were investigated using negative binomial regression models; incidence rate ratios (IRRs) and 95% confidence intervals (CIs) are reported. Consultation year was fitted as a fixed effect, GP practice as a random effect, and adjusted practice list size (per level of covariate where appropriate) as the offset. Fixed effects for each categorical variable (age, sex, ethnicity, IMD quintile, mental health status, and shielding status), along with the interaction between each covariate and consultation year, were separately fitted to the model; interaction p-values are presented and results only presented separately for each level of a covariate if p<0.05. Model validity was checked using standard methods; outliers which disrupted model fit were removed.
To investigate consultation provision (i.e. face-to-face vs. remote consulting), we modelled proportions of consultations with potential cancer indicators. Separate models were fitted for GPs and nurses/paramedics consultations. Negative binomial regression models were fitted with number of consultations as the outcome, a fixed effect for consultation provision and an interaction with year, GP practice as a random effect, and total numbers of consultations per practice per level of consultation provision as the offset.
For individual potential cancer indicators, unadjusted IRRs comparing April-July 2020 to 2019 are presented to help interpretation, but due to large numbers of indicators and issues with multiple testing, no modelling was performed.
Stata 15.1 was used for all data management and analyses.