Continuity of care in general practice at cancer diagnosis (COOC-GP study): a national cohort study of 2853 patients

Study design

This was a post-hoc analysis of a prospective cohort study based on data from the Observatoire de la Médecine Générale (Observatory of General Practice, http://omg.sfmg.org). From 1993 until 2012, 129 French GPs prospectively included every patient consultation in routine practice and provided standardised data on patient characteristics and diagnoses. The GPs are from 11 out of 22 different regions. The participants in this network were largely representative of the French GP population, although a comparison with data from the Ministry of Health showed that doctors working in rural areas were under-represented.²⁰

Using this cohort, the Continuité des Soins en Médecine Générale au Diagnostic de Cancer (Continuity of Care in Cancer Patients in General Practice, COOC-GP) study focused on adult patients diagnosed with an incident cancer between 1 January 2000 and 31 December 2010. The date of the first consultation in which the diagnosis was recorded was used as a proxy for the date of the cancer diagnosis. The period assessed for each patient was from 18 months before diagnosis until 12 months after diagnosis. The baseline period was considered to be the 1-year period from 18 months until 6 months before the cancer diagnosis; this was done to avoid the known increase in general practice care consumption related to the diagnosis.¹⁹ Patients were excluded if their usual consultations with a GP were too infrequent — this was defined as there being no consultation during the 1-year baseline period or during the year before diagnosis.

Data

The collected data included patient and GP demographic characteristics, a description of the first consultation with the diagnosis of cancer mentioned in the GP medical record, and dates of other consultations. The tumour site was classified by group:

• breast;

• genital, female;

• genital, male;

• digestive system (the reference group);

• skin;

• respiratory system;

• lymphoma/leukaemia; and

• other locations.

Patients were described as ‘living in rural areas’ or not (because of the difference in access to secondary care); the threshold for rural areas was defined as <2000 inhabitants.

A baseline comorbid condition score was generated using the comorbidities of the Charlson comorbidity²¹ index documented before the cancer diagnosis, and combined using modified weights (0 = no comorbidity, 29 = maximum score).²² Bannay’s modified Charlson index scores 12 comorbidities with the following weights:

• congestive heart failure = 2;

• peripheral vascular disease = 1;

• cerebrovascular disease = 1;

• dementia = 2;

• chronic pulmonary disease = 1;

• mild liver disease = 2;

• moderate or severe renal disease = 1;

• hemiplegia = 2;

• any tumour (including leukaemia and lymphoma) = 2;

• moderate or severe liver disease = 3;

• metastatic solid tumour = 11; and

• HIV/AIDS = 1.²²

The maximum score of 29 is only theoretical.

COC was estimated using the maximal interval, in days, between two consultations to reflect the patient’s usual consultation regularity. It was coded as a categorical variable:

• frequent: 0–91 days (0–3 months);

• regular: 92–182 days (>3–6 months); and

• occasional: ≥183 days (>6 months).

These categories were chosen taking into account French regulatory drug packaging (28–30 days or 84–90 days), and there being a maximum 6-month prescription duration for patients with chronic conditions and a mean of seven medical consultations per patient per year in France.²³

The main outcome was loss of COC, defined as a longer interval between consultations with the same GP in the measurement periods than in the baseline period. Patients who died were excluded. In the sensitivity analysis, this outcome was tested with a 30-day and a 20% margin above the longer interval; all patients lost to follow-up were considered a loss of COC. These margins were chosen by the authors to account for there being a 1-month maximum duration for prescriptions given at hospital discharge or for a pharmacy exceptionally dispensing without a new prescription.

The ongoing relationship between the patient and GP was also determined by the number of consultations per patient per trimester, and the proportion of patients seen at least once per trimester. COC was measured during the trimester before cancer diagnosis to observe the diagnostic procedures period, and the year after to observe the cancer management and treatment period. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) recommendations were adhered to.

Statistical analysis

Age was coded into four groups:

• 18–49 years;

• 50–64 years;

• 65–74 years; and

• ≥75 years.

Each patient’s comorbidity score was coded as a three-class variable:

• no metastasis and no comorbidity: score 2;

• no metastasis with comorbidities: score 3–12;

• metastatic cancer: score ≥13.

The paired Student’s t-test and Wilcoxon matched-pairs signed-rank test were used to compare the maximal interval between consultations and the distributions of the frequencies of consultations between the measurement period and baseline period respectively. Differences between groups with frequent, regular, or occasional consultations were tested using a χ² test or analysis of variance as appropriate, and the Cuzick test, a non-parametric test for trend across ordered groups.²⁴

Factors associated with loss of COC were analysed separately during the two study periods with univariate and multivariate logistic regression models. Variables with a P-value of <0.2 at univariate analysis were introduced in the multivariate analysis after correlations among variables were tested with Cramér’s V and Pearson’s R² statistics; confounding was also tested. The baseline frequency of consultations was used as a continuous variable (that is, number of days between two consultations per patient) and expressed in months. GPs’ variability was tested in hierarchical mixed models. The discrimination abilities of the multivariate model were checked by area under the ROC curve (AUC) analysis. For all analyses, Stata version 14.2 was used and statistical significance was set at P<0.05.