Frequency and impact of medication reviews for people aged 65 years or above in UK primary care: an observational study using electronic health records

This was an observational study using electronic health records (EHR) to capture medication reviews recorded from 01 January 2019 to 31 December 2019. We focused on reviews occurring in 2019 as the last complete year before the introduction of the new Structured Medication Review guidelines [13]. We used data from the UK’s Clinical Practice Research Datalink (CPRD GOLD), a database of anonymised routinely collected primary care EHR. CPRD patients are representative of the UK population in terms of age and sex, and in 2013 it was estimated that 6.9% of the total UK population were included [15]. The data are captured as part of routine healthcare delivery and administration, and include demographic and lifestyle characteristics, symptoms and diagnoses, test results, prescriptions issued in primary care, and other clinical events. Clinical information is coded using Read codes (v2). We used the CPRD GOLD dataset (May 2022 build) linked with person-level and practice-level deprivation data (Townsend Score quintile [16], linkage based on post code).

The study included people aged 65 years and over in 2019 with at least one year of ‘up-to-standard’ [15] (a marker of data quality) follow-up in CPRD before 01 January 2019. Individual patient follow-up was from 01 January 2019 to the earliest of death, leaving the practice, or last practice data collection date. Due to small numbers (n=13), we excluded people with missing/unclassified sex. For the analysis looking at numbers of people having a medication review, we excluded people with no ongoing prescriptions at baseline (defined below) and censored follow-up at first medication review or 31 December 2019. For the analyses comparing types and numbers of medicines prescribed before and after a medication review, we included only people who had a medication review in 2019 and at least three months’ follow-up after the review. When comparing numbers of medicines, we also excluded people with no prescriptions in the three months prior to their medication review, and with an extreme change in count (more than three standard deviations (SDs) from the mean – see later).

Medication reviews recorded during 2019 were identified using a pre-defined list of Read codes (see Additional File 1, table S1.1; also available online – see data sharing statement). The code list was developed by searching the code dictionary for relevant terms. The final list was approved by the research team, which includes practicing general practitioners (AJA, BI). The first medication review recorded in 2019 was the event of interest and any subsequent reviews are not included in the study. Medication reviews were classified according to the consultation type (face-to-face, telephone, other) and staff role (pharmacist, GP, nurse, admin, other) associated with each record (details in Additional File 1). As a sensitivity analysis, we used a conservative definition of medication reviews including only reviews recorded during face-to-face or telephone consultations, and excluding reviews for specific medicines/conditions, identified using Read codes (see Additional File 1, table S1.1).

We derived the numbers and types of medicines prescribed at baseline, and in the three months before and three months after a medication review. Medicines were classified by drug substance and formulation, British National Formulary (BNF) [17] paragraph, and BNF chapter, and prescriptions were classified according to whether they were issued as a repeat prescription (yes/no). We excluded products in BNF chapters 14 (Immunological Products and Vaccines) and 15 (Anaesthesia), and non-medicinal products such as dressings, devices, and garments. The number of prescribed medicines (referred to as the ‘prescription count’) was defined as the number of unique medicines with prescriptions that overlapped for at least one day. The duration of each prescription was estimated using a similar approach to Pye et al. [18], using information about quantity prescribed and daily dose, and making different assumptions based on drug classes and formulations. Products with the same drug substance but different formulations were classified separately. The prescription count was categorised as 0, 1, 2-4, 5-9, 10-14, 15-20, and 20+ prescribed medicines. Further details are given in Additional File 1 and the data processing code is available online (see data sharing statement).

Baseline variables were defined with respect to 01 January 2019 and were modelled as binary or categorical variables, except age (continuous, and categorised into 10-year bands). These variables were: age in 2019, sex, ethnicity, deprivation (quintile of Townsend Score [16], person-level if available, otherwise practice-level), practice region (grouped as Scotland, Wales, Northern Ireland, London, rest of England), smoking status, alcohol intake, body mass index (BMI), having a medication review in the previous year, whether the person was living in a care home, diagnoses mentioned in the Structured Medication Review guidelines [13], the Quality and Outcomes Framework [19], or the New Medicine Service [20], and baseline medicines commonly associated with medication errors [21]. The diagnoses included were: atrial fibrillation, cancer, chronic kidney disease, chronic obstructive pulmonary disease (COPD), coronary heart disease, dementia, depression, anxiety, diabetes, epilepsy, heart failure, hypertension, hypothyroidism, learning disability, mental health disorder (schizophrenia, bipolar affective disorder, or other psychosis), obesity, osteoporosis, being on the palliative care pathway, peripheral arterial disease, rheumatoid arthritis, stroke or transient ischaemic attack (TIA), asthma, dyslipidaemia, gout, glaucoma, Parkinson’s disease, benign prostatic hyperplasia, urinary incontinence or retention, mobility problems, thrombosis or thrombophilia, and frailty (severe frailty, recent fall, recent fracture). Where available, we used existing Read code lists to define these variables. These were sourced from the ClinicalCodes repository [22], the HDR-UK Phenotype Library [23, 24], OpenCodelists [25], QOF business rules [26], and individual papers (see Additional File 1). Baseline medicines were considered present if prescribed on or in the six-months before 01 January 2019 and included: non-steroidal anti-inflammatory drugs (NSAIDs), aspirin/antiplatelet medicines, renin-angiotensin system drugs, diuretics, opioids including combination analgesics, antidepressants, antipsychotics, bisphosphonates, benzodiazepines and z-drugs, gabapentinoids, inhaled long-acting beta-antagonists and corticosteroids, lithium, and anticholinergic medicines. Finally, we created indicator variables for medicines with ongoing prescriptions at baseline (i.e., with start and stop dates spanning 01 January 2019), at BNF-chapter level. Further details about how we defined the variables in this paragraph are given in Additional File 1.

We summarised the baseline characteristics of the population of people with at least one ongoing prescription on 01 January 2019. We assessed the association between baseline characteristics and having a medication review in 2019 using Cox regression in order to take follow-up time into account. We performed unadjusted models, age-sex adjusted models, and multivariable models including demographic and clinical characteristics, and tested the proportional hazards assumption by examining log-log plots of survival and comparing observed and predicted Kaplan Meier survival plots.

When comparing types of medicines, we included only medicines flagged as repeat prescriptions to reduce noise from short-course medicines (e.g., anti-infective medication), which were overrepresented when considering medicines prescribed only before or only after a review (see Additional File 3 Figure S3.4 for results including all medicines). For simplicity and readability, we refer to medicines only prescribed before a review as ‘stopped’, medicines only prescribed after a review as ‘started’, and medicines prescribed before and after a review as ‘continued’. We ranked the twenty medicines most frequently stopped and/or started after a medication review, and counted the numbers of people who stopped, started, or continued each medicine. These results are presented graphically.

To assess the change in maximum prescription count, we calculated the difference between maximum prescription count after the review and maximum prescription count before the review. We then excluded people with a change in count more than three SDs from the mean, reducing skew when breaking down results by original prescription count. After confirming the change in prescription count had a Normal distribution, overall (Additional File 2 Figure S2.1) and for the different levels of the categorical variables, we used multivariable-adjusted linear regression to test the association between change in prescription count and demographic characteristics, prescription count before the review, consultation type, and staff role.

We repeated all analyses using the conservative definition of medication reviews and assessed change in type and number of medicines prescribed before and after a review using different time windows: 6 months before and after, 3 months before vs 1-4 months after (in order to account for any gradual changes, such as tapering medicines before they were stopped), and 1 month before and after.

The study team includes three patient and participant involvement representatives (DB, LG, DW) who helped define the study aims and who have provided insights on our study design, methods, and interpretation.

Figure 1 shows the number of people included in each part of the analysis. The original dataset contained 722,404 males and females aged 65 years and over with at least one year of up-to-standard follow-up prior to 01 January 2019 (baseline). After excluding 130,678 (18.1%) people with no ongoing prescriptions on 01 January 2019, 591,726 people remained in the study population for the first part of the analysis. The median age in this group was 74 years (interquartile range 70 to 81 years), and 54.5% were female. Key baseline characteristics are presented in Table 1, and further clinical details are presented in Additional File 2 Table S2.1.

Out of the 591,726 people in this part of the analysis, 305,526 (51.6%) had at least one medication review in 2019. Of these, 73.2% had a single medication review. Considering only the first recorded medication reviews, 45.5% were recorded during face-to-face or telephone consultations and 67.2% were recorded by GPs (Table 2).

Results of unadjusted, age-sex-adjusted, and multivariable-adjusted Cox regression models are shown in Table 3 (full results for the models are given in Additional File 2 Table S2.2). In the unadjusted and age-sex adjusted models, most baseline diagnoses and prescribed medicines were associated with an increased likelihood of having a medication review (Additional File 2 Table S2.2). For diagnoses and prescribed medicines, the largest associations with having a medication review were for atrial fibrillation (age-sex adjusted hazard ratio (HR) 1.29, 95% CI 1.25 to 1.33) and antipsychotics (age-sex adjusted HR 1.38, 95% CI 1.32 to 1.44), respectively. In the fully adjusted model, living in a care home (HR 1.51, 95% CI 1.40 to 1.62), medication review in the previous year (HR 1.83, 95% CI 1.69 to 1.98), and increasing prescription count at baseline (HR for 5-9 medicines vs 1 medicine 1.41, 95% CI 1.37 to 1.46) were most strongly associated with having a review (Table 3). There was only a weak association between older age and having a medication review (HR for 75-84 vs 65-74 years 1.03, 95% CI 1.02 to 1.05), and no association for sex (HR 1.00, 95% CI 0.99 to 1.01 for male vs female). There were differences according to UK region, with people at practices in London (HR 0.62, 95% CI 0.47 to 0.81) and the rest of England (HR 0.77, 95% CI 0.65 to 0.91) less likely to have a recorded medication review than Scotland (the largest group in this case).

In a sensitivity analysis using the conservative definition of medication reviews, 142,576 people (24.1% of the eligible population) had a medication review in 2019. The Cox regression results were broadly similar to the main analysis (Additional File 2 Table S2.3). One notable difference was in the association with practice region: compared to Scotland, people at practices in Wales were more likely to have a recorded medication review (HR 1.29, 95% CI 1.10 to 1.50) and people at practices in Northern Ireland less likely (HR 0.66, 95% CI 0.49 to 0.90). There was no significant difference between people at practices in London or the rest of England and people at practices in Scotland.

As shown in Fig. 1, 309,207 people who had a medication review in 2019 and at least three months of follow-up after their medication review were included in this part of the analysis. Characteristics of these people are summarised in Additional File 2, Table S2.4.

Figure 2 shows the twenty medicines most frequently stopped or started in the three months after compared to three months before a medication review (see Additional File 2 Figure S2.5 for the corresponding statistics). This figure is restricted to tablets, but a version including all formulations is provided in Additional File 3. For the medicines shown, most people had prescriptions both before and after a medication review. From this list, four medicines were slightly more frequently stopped than started (aspirin, simvastatin, bendroflumethiazide, and warfarin). The other medicines shown were more often started than stopped. Some medicines made the list due to being prescribed to large numbers of people at any time (e.g., atorvastatin, omeprazole, simvastatin). Other medicines were prescribed to fewer people in total, but a higher proportion of people started or stopped them compared to other medicines. These included vitamins/supplements (e.g., colecalciferol, folic acid, ferrous fumarate), alendronic acid, ranitidine, apixaban, and warfarin. Additional figures showing results for sensitivity analyses, plus breakdowns at BNF chapter level and for named drug groups, are provided in Additional File 3. A sensitivity analysis using the strict definition of medication reviews had very similar results to the main analysis (Additional Figure 3 S3.12). Similarly, increasing the time window around the reviews did not have a big impact on the findings. With the shorter window (one month before and after the review), the proportions of people who stopped or started individual medicines were notably larger (likely reflecting refill frequencies slightly longer than one month).

When comparing the change in maximum prescription count, we further excluded people with no prescriptions in the three months before the review (n=11,949), and people with a change in maximum prescription count more than three standard deviations from the mean (n=5311), meaning 291,947 people remained in the analysis. The median (interquartile range) maximum prescription count was the same before and after a review (5, 3 to 8 medicines). The mean difference in maximum prescription count after compared with before a review was 0.13 (95% CI 0.12 to 0.14, P<0.001), indicating a small but statistically significant increase in count. Categorised prescription count before/after a review is summarised in Additional File 2, Table S2.6. Table 4 shows the results of multivariable linear regression adjusting for demographic characteristics and features of the medication review. Most factors had very small associations with change in prescription count. On average, people with a maximum prescription count of 10 or more medicines in the three months before a review had a lower count after the review (represented graphically in Additional File 2 Figure S2.2).

A sensitivity analysis using the strict definition of medication reviews, which included 136,139 people, found similar results, with a mean difference in maximum prescription count after compared with before the review of 0.15 (95% CI 0.14 to 0.16, P<0.001). Increasing the time window around the medication review did not affect the results (six months pre- and post-review, mean change 0.17 (95% CI 0.16 to 0.17, P<0.001), three months pre- and one to four months post-review, mean change 0.09 (95% CI 0.08 to 0.10, P<0.001)). However, with a shorter time window (one month pre- and post-review), the mean change in prescription count was -0.48 (95% CI -0.54 to -0.42, P<0.001), indicating a decrease in count on average.