Study design
For the purpose of this retrospective study, primary care utilisation of known diabetes patients was assessed in 2008 by developing profiles of diabetes-related healthcare utilisation and of total health-care utilisation separately. Data from the Netherlands Information Network of General Practice (LINH) were used, which is a representative sample of GP-practices in the Netherlands that provide routinely recorded data from their electronic medical records (EMRs) of all patients listed in their practice. The Dutch healthcare system is very useful for analysing longitudinal data. All Dutch inhabitants are obligatory listed in a general practice and the GP acts as gatekeeper for specialized health care. Therefore, the EMR kept by the GP is the most complete record. The LINH-database holds longitudinal data on morbidity, drug prescriptions and referrals of approximately 90 GP-practices and 350,000 listed patients [15]. The network is a dynamic pool of practices, with each year some minor changes in the composition of practices. Diagnoses are coded by the GPs using the International Classification of Primary Care (ICPC) [16]. LINH is registered with the Dutch Data Protection Authority; data are handled according to the data protection guidelines of the authority. According to the Dutch legislation, ethical approval is not required for observational studies.
For our analyses, we used data from practices that a) participated in both 2007 (for determining known diabetes patients) and 2008 and b) provided recorded year-round data for consultations, prescriptions, and morbidity and referral records in 2008.
Patients were selected for this study if 1) they had consulted their GP for type II diabetes at least once in 2007 and 2) were registered with the practice during the whole year in 2008 and 3) were 18 or older. Type II diabetes patients were selected on the basis of a recorded ICPC-code T90. GPs participating in LINH do generally not record on ICPC sub codes (T90.1 or T90.2), and therefore we could not distinguish between type I and type II diabetes patients on the basis of ICPC-codes. For the purpose of this study, type I diabetes patients were excluded on the basis of having received a prescription of insulin (ATC-code A10A), but not any oral anti-diabetic medication (ATC-code A10B) [14, 17]. In total, data of 48 GP-practices and 6,721 type II diabetes patients were included. Reasons for exclusion were 1) incomplete data on consultations (16% of practices), 2) incomplete data on prescriptions (28%) and/or 3) incomplete data on referrals (44%).Overall, these GP-practices were representative of the Dutch GP-practices with respect to degree of urbanisation and region, but not with respect to practice type (overrepresentation of group practices or health centres, underrepresentation of single handed practices).
Healthcare utilisation in primary care
Healthcare utilisation of subjects consisted of contacts with general practice, drug prescriptions and referrals to allied healthcare: primary healthcare and medication. Healthcare utilisation was regarded as diabetes-related, if the care provided was mentioned in the multidisciplinary healthcare standard of the Dutch Diabetes Federation (NDF: “Nederlandse Diabetes Federatie”) [18]. We only took into account healthcare utilisation for which an ICPC code was recorded by the GP. Additional file 1 shows the definition of diabetes related healthcare for known type II diabetes patients with the corresponding ICPC-codes.
Contacts in general practice were derived from claims data in the EMR in which a distinction was made between consultations in the practice, home visits and telephone consultations with GPs and primary care nurses. The numbers of consultations, home visits and telephone consultations with both GPs and primary care nurses separately, were included in the analyses.
Drug prescriptions were coded using the ATC classification system. Prescriptions mentioned in de NDF-guideline (see Additional file 1) as well as prescriptions related to the diabetes-related health problems were included in the analyses of diabetes-related profiles. A list of included drug prescriptions is provided in Additional file 2. The number of different drug prescriptions based on ATC codes at the 4 digit level was included in the analyses. If a prescription could be captured with an ATC code at the 5 digit level, the ATC code at the 5 digit level was included.
Healthcare utilisation of patients with allied healthcare providers was estimated on the basis of GP referrals. Referrals for diabetes-related health problems to a physiotherapist or exercise therapist, dietician and podiatrist were included in the analyses.
Patient and disease characteristics
Patient and disease characteristics included in this study were age (categorised), gender, urbanisation (categorised), diabetes medication type (no treatment, oral treatment or oral treatment and insulin) and comorbidity; all data were derived from the EMR. Comorbidity was divided into several categories of related and unrelated comorbidity according to Struijs et al. (2006) [19]. The following comorbid diseases were regarded as diabetes-related: (with ICPC-codes): heart diseases (K74-K76), stroke (K90), retinopathy (F83), nephropathy (U99) and diabetic foot (K99.06 and N94; deviated from Struijs et al.). Non-related comorbidity included depression (P76), lung diseases (R91, R95, and R96), musculoskeletal diseases (L01-L03, L08, L13, L15, L84, L86, and L89-91), neurological diseases (N86-N88) and cancer (B74, D74, D75, D77, R84, S77, X76, Y77).
Statistical analyses
To analyse the different profiles, Latent Class Analyses (LCA) were performed to identify distinct classes of patients with specific combinations of healthcare utilisation. LCA is a type of cluster analysis used to group patients into k number of unique (otherwise unobserved) categories, where, within each category patients are most similar to each other regarding their healthcare utilisation, and between the categories patients are most different [20–22]. To find the optimal number of categories, a 2–6 class solution was modelled and output was assessed and compared according to a stepwise approach described elsewhere [20, 22]. To determine the final solution several model fit indicators were used [23]. The Bayesian Information Criterion (BIC) (where a lower BIC indicates a better fit) and posterior probabilities (where probabilities close to 1 indicates a better classification and posterior probabilities at least 0.8 are advised [24, 25]) were used as model fit indicators. Also, we assessed the usefulness and clinical interpretation of each solution. The usefulness was assessed by considering the solutions based on the number of people in each class (hereby rejecting solutions with small groups: minimum N = 200). Mplus was used to perform LCA because within Mplus, LCA can adequately cluster a combination of both categorical (also binary variables) and count data [26]. LCA was conducted for both diabetes-related healthcare utilisation and total healthcare utilisation separately. Each profile was given a label resembling their healthcare utilization. Subsequently, a predictive model was made using multilevel multinomial regression analyses (patients nested in practices) for the diabetes-related healthcare utilisation profiles. In this analysis, it was assessed whether patient and disease characteristics were associated to profile membership. Analyses were performed using STATA, Mplus and MLwiN.