Distinguishing appointment patterning in primary care: an unsupervised machine learning approach

Problem

Since the COVID-19 pandemic, primary care has pivoted towards telephone consultations and adopted same-day care models to accommodate increased patient demand. However, there is sparse evidence on the implications of such same-day triage systems for primary care equity, efficiency, and accessibility. The heterogeneity in implementing these systems, which vary across practices by appointment type, consultation modality, and healthcare workforce utilisation, complicates the categorisation of GP appointment systems using administrative data. This study addresses the challenge of distinguishing between same-day and other appointment systems within routine datasets using unsupervised machine learning.

Approach

We used the Appointments in General Practice dataset from NHS England from October 2023 . This publishes appointment data crosstabulations at practice level. We included practices with greater than 90% of appointments mapped to an NHS Digital harmonised appointment type. Eight variables were derived a priori, deemed likely to differentiate practice appointment systems: the proportion of total appointments that were booked more than one week in advance; with a GP; GP telephone consultations; coded as clinical triage; same-day GP appointments; telephone consultations; acute telephone consultations; appointment rate per 1000 patients .K-means clustering with 25 randomisations and up to 500 iterations was used to group practices into clusters. The optimal number of clusters was determined by silhouette width. Differences in sociodemographic characteristics between the two clusters were described using Census 2021 data and Quality Outcomes Framework data for 2022-2023.

Findings

Of 6,290 practices listed in the dataset, 4,080 (65%) had adequate data quality for analysis.K-means clustering yielded two distinct clusters. Cluster 1 (“Acute care practices”, n=1539) comprised practices tending towards same day care, delivered by telephone by GPs with triage support. Cluster 2 (“Routine care practices”, n=2539) was characterised by a greater proportion of appointments over one week in advance (40% of total appointments vs 27% for acute care practices), greater use of non-GP appointments (proportion of appointments with a GP 40% vs 54%) and fewer telephone appointments (16% versus 34%).Routine care practices tended to be in rural settings (16%, vs 10% of acute care practices) and were less likely to be in London (11% versus 30%). Acute care practice patients were younger on average (mean age 39.8 vs 41.8; percentage over 65 years 16% vs 19%) and less likely to identify in a White ethnic group (71% White versus 84% in routine care practices). Acute care practices are also more likely to be situated in the most and second-most deprived quintiles of England (55% vs 50%).

Consequences

This study reveals two divergent approaches to UK primary care appointment access. Acute care practices tend to have younger and more diverse and deprived populations. Further work will describe any implications for cardiovascular screening, long-term conditions management and patient-recorded quality of care.