Study design, data source and participants

In NWL, local health care providers have worked together to produce the Whole Systems Integrated Care (WSIC) dataset, comprising linked electronic health and social care records covering 2.4 million people registered to GP practices in NWL [23]. We conducted a retrospective cohort study using de-identified WSIC data extracted on 1^st November 2023, with study period 1^st Jan 2022 to 31^st Dec 2022. The authors did not have access to information that could identify individual participants during or after data collection. Participants were 18 years or older, and registered with a GP in NWL, at the start of this period. We linked demographic and disease condition data for every adult to their prescription records for the study period.

Variables

We defined polypharmacy to be five or more distinct regular medications prescribed to a patient during the study period. To count distinct medications, we used SNOMED CT disposition categories [24]. These dispositions correspond with drug functions within the body. For example, statins form a disposition called “3-hydroxy-3-methylglutaryl-coenzyme A reductase inhibitor”. The link from a prescribed medication to a disposition is made via the medication’s active ingredients, the chemical substance(s) it is formed from. This is illustrated for statins in Table 1. Some active ingredients in the SNOMED CT ontology do not have an assigned disposition. A general practitioner and a pharmacist independently reviewed these active ingredients, and agreed on one of two classifications for each. First, ingredients not relevant for polypharmacy (e.g. mineral oil, alginic acid, virus antigens) were excluded. Others relevant for polypharmacy (e.g. metformin, paracetamol, laxatives), were assigned to an existing SNOMED CT disposition if possible, or a newly created disposition if not. A full list of these assignments is provided in S4 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Disposition medications: Statins.

List of medications prescribed to more than 1000 patients during the study period, associated to the Statin disposition in SNOMED CT with their full list of active ingredients and the number of patients who were prescribed that medication regularly within the study period. Full list available in S5 Table.

https://doi.org/10.1371/journal.pone.0308624.t001

Prescribed items were excluded if they did not contain an active ingredient (e.g. casts and crutches). We also excluded non-oral medications (e.g. injections, gels, ointments; full list in S6 Table), and medications with seven or more active ingredients (all multivitamin medications) which would skew counts. A small number of medications were excluded as not relevant for another reason (S6 Table).

A disposition was deemed to be “regular” for a patient if it was prescribed three or more times for that patient during the study period. Where there were multiple prescriptions of the same disposition for the same patient within 10 days, these were only counted once. The number of regular dispositions identified per patient was used to categorise patients as having 0, 1 to 4, or 5 or more regular medications, the latter defined to be polypharmacy.

Drug costs were taken from the NHS Business Services Authority Dictionary of Medicines and Devices [25]. Recorded drug costs were presented as totals and averages per patient over demographic categories.

We extracted demographic variables for each individual: age, gender, ethnicity, a code for the GP practice the individual was registered with, and Index of Multiple Deprivation (IMD) quintile [26]. IMD is a statistic of relative socio-economic deprivation derived for small areas in England, UK. We extracted information on components of the electronic frailty index (eFI), a tool quantifying frailty in terms of deficits an individual has, as documented in the EHR. The eFI was designed and validated with patients aged 65–95 [27] and contains 36 components including Arthritis, Diabetes, Dizziness, Weight loss and anorexia, and Hypertension. Polypharmacy is also a component of the eFI defined as prescription of ≥5 prescribed medications, using chapters 1–15 of the British National Formulary [27], this was excluded from the analysis in favour of the definition described above.

Statistical analysis

The unit of observation was the patient. We described demographics stratified by number of regular medications. We used a mixed-effects multivariable logistic regression model to determine factors associated with polypharmacy, modelling variation between GP practices through a random intercept. Age, gender, ethnicity, deprivation quintile and number of eFI components were fixed effects. Missing data for ethnicity were treated as a separate level. Regression diagnostics were performed including examination of residuals, a check for multicollinearity and assessment of the Akaike Information Criterion. Results were reported as odds ratios (ORs) with 95% confidence intervals (CIs), and p-values at a statistical significance level of 0.05. We included interaction terms to examine any modification of the effect of eFI by gender or ethnicity. For all pairs of regular medications, we calculated the proportion of patients prescribed both, and plotted the results as a heatmap.

We performed all statistical analyses using R version 3.6.1.

The NHS Health Research Authority West Midlands—Solihull Research Ethics Committee provided written ethical approval covering this study under the Discover Research Platform (REC: 18/WM/0323, IRAS: 252449), waiving the need for informed consent since the study was retrospective and used de-identified data only. Data access for this study was approved by the North West London Data Access Committee (ID-271-2). All methods within this study were carried out in accordance with the guidelines and regulations of the ethics and study approval.

Summary

This study of 1.7 million adults found overall prevalence of polypharmacy is 9.4%. Older age, greater frailty, and living in an area of higher socio-economic deprivation were independently associated with polypharmacy. Having adjusted for frailty, male gender was associated with polypharmacy. With average frailty, Black people were less likely than other ethnicities to be on polypharmacy. However, the effect of increasing frailty was greater for all non-White ethnicities.

Our results on direct cost of regular medications implicated in polypharmacy are the first derived from electronic health records in the UK, and demonstrate the considerable resources involved. Notably, the cost per patient on polypharmacy is greater than expected simply based on the number of regular medications prescribed, suggesting that patients on polypharmacy may be prescribed more expensive medication on average than those on fewer regular medications.

We developed a new method for identifying polypharmacy in primary care data, using the SNOMED CT ontology. This method enabled us to identify groups of medications by function in the body, count regularly prescribed medications and identify common combinations of medications prescribed for the same patient. The most common combinations included cardiovascular medications, including statins with calcium channel blockers or proton pump inhibitors. Combinations associated with a high risk of adverse side effects, such as angiotensin II receptor blockers (ARBs) with angiotensin-converting enzyme (ACE) inhibitors, occur with low frequency, but were still prescribed for some patients.

Strengths and limitations

The WSIC dataset provided high coverage of the population and full coverage of primary care prescriptions. This minimised selection bias compared with a sampling approach. The population of NWL is large, diverse, and broadly representative of the demographics of the wider UK population, strengthening the external validity of the study [23]. Ethnicity was well documented in the study population, enabling better understanding of potential health inequalities.

Our method for identifying polypharmacy ensured medications only differing in dosage, and medications with the same function, were not double counted (S1 Box). For example, switching a patient from atorvastatin to simvastatin part way through the study period would not add to their count of regular medications, since both have the same disposition. We deliberately chose to count drugs by their function rather than by body system, as in the BNF, for two reasons. First, primary care EHR databases do not systematically record the indication for each prescription. Therefore it is not possible to reliably map each prescription onto the problem or disease, and hence body system, it was prescribed for. Second, this approach of counting distinct drug functions in the body has the advantage of highlighting cases where a large number of fundamentally different drugs are prescribed for the same patient, better reflecting complex polypharmacy. The approach is therefore well suited to describing polypharmacy from the important perspective of potential for drug-drug interactions. In contrast, this method is not suitable for identifying issues of therapeutic duplication involving two or more drugs in the same disposition group. This limitation is shared by other approaches using groups of medications to count polypharmacy, for example BNF chapters. The method is widely applicable to any SNOMED CT coded prescription data. Our mixed-effects multiple regression model enabled a precise understanding of independent associations between each variable and polypharmacy, and this is the first comparable study to separate out interactions between demographics and frailty.

This study was limited by the data, which were routinely collected and therefore dependent on coding by primary care teams. We had data on prescription but not dispensing of medication. Therefore, whilst we were able to quantify and describe patterns in polypharmacy as intentionally prescribed by a clinician, we were not able to characterise to what extent prescriptions were dispensed by a pharmacist or collected or used by patients. Patterns of polypharmacy as ingested or otherwise used by patients may differ from prescription patterns such as those identified in this study, for example where there are differences in adherence by demographics [29]. The eFI is a well-established and validated measure of frailty [27]. A 2023 systematic review found that eFI was the best-performing validated electronic measure of frailty, although the review only included one study using eFI [30]. However well the eFI performs as a measure of frailty, like all such measures it will not capture all aspects of frailty relevant to polypharmacy. Since this study is observational and does not establish causal explanations for the patterns identified, this is not a major limitation, but it is important to note that some of the variation associated with other variables may be partly explained by unmeasured frailty.

The NHS underwent radical changes during the height of the COVID-19 pandemic, for example many more GP consultations were conducted remotely [31]. All major public health measures, such as lockdown restrictions, were lifted by summer 2021 in England. However, it is possible that some differences in prescribing during this time persisted into the study period. However, given that the study period commenced six months after the lifting of restrictions, it is likely that the findings of this study reflect post-lockdown prescribing practices in the NHS and hence are relevant in informing policy and practice.

Comparison with existing literature

Two previous studies from Scotland found higher rates of polypharmacy (five or more medications) than were identified in this study, Payne et al. finding approximately 17% prevalence in 2006, and Guthrie et al. 21% prevalence counting distinct chemical substances, and 7% counting BNF chapters, by 2010 [5, 16]. These studies were set in different populations, and used different methods, with Payne et al. using a simple count of prescriptions, rather than classifying drugs into categories, and Guthrie et al. using 355 subgroups of an amended British National Formulary (BNF) classification, splitting BNF chapters containing multiple drugs considered distinct and often co-prescribed. Similar to Payne et al. [5], an NHS report on polypharmacy counted by distinct medications and presented that 15% of people in England in 2019 were on five or more medications [19]. A study using electronic health records from the Spanish National Health System, using the Anatomical Therapeutic Chemical active substance classification, found a prevalence of 8.9% by 2015, more similar to the prevalence identified in this study [18]. All studies used different methods to characterise regularity of prescription in defining polypharmacy, impacting prevalence figures.

Patterns in polypharmacy recorded in the literature were broadly consistent with our findings, with greater prevalence associated with increasing age and higher levels of socio-economic deprivation. Payne et al. found increasing prevalence of polypharmacy with higher numbers of clinical conditions, consistent with our findings in relation to frailty [5]. Our finding that unadjusted prevalence of polypharmacy was greater among women than men was reversed on adjusting for number of frailty components. This reflects higher average levels of frailty among women compared with men, coupled with the strong association of frailty with polypharmacy. Payne et al. similarly found no statistically significant difference between men and women when adjusting for number of clinical conditions, where women were more likely to be prescribed more medications in univariate analysis [5]. Our finding of increased odds of polypharmacy with increasing socio-economic deprivation is consistent with previous studies [5, 16]. This consistency with existing literature supports the generalisability of the trends and patterns in our findings for populations and health systems that are similar to the National Health Service in England (NHS), and particularly Northwest London. There will of course be variation in prescribing practices both within and between such systems. It is likely these findings will not generalise to populations and health systems that differ significantly from the NHS, for example where healthcare policy, population needs, and available medications differ.

Implications for research and practice

Our study shows that one in ten adults are on polypharmacy, and that whilst in many cases this may be beneficial, for some patients potentially problematic drug combinations may be prescribed. This supports the use of interventions to reduce problematic polypharmacy, such as structured medication review (SMR), as recommended by NHS England [32]. Such interventions create space for shared decision making between patients and clinicians, with potential to improve drug adherence, improve health outcomes, and reduce waste [33]. There is limited evidence on roll out of SMR and its effectiveness at population level [34]. Future research should quantify uptake of SMRs in practice, and examine their effectiveness at scale.

Our results provide evidence not only of well-known patterns of polypharmacy such as increasing prevalence with age, frailty, and lower socio-economic deprivation, but also of new associations, particularly in relation to interactions between frailty, and gender and ethnicity. This is the first study of large electronic health record data to examine the relationship between ethnicity and polypharmacy. Our findings indicate that the burden of polypharmacy becomes greater with increasing frailty to a larger extent among people of Black, Asian and Other ethnicities than it does among those of White ethnicity.

This differential effect of frailty for different genders and ethnicities suggests these interactions are important in research and in informing practice. A UK-based study of patients with dementia found differences in prescribing patterns by ethnicity, and another found variation in prescription of anti-osteoporosis drugs associated with income deprivation [35, 36]. Both are examples of differences in prescribing associated with demographics among patient groups who have higher than average frailty, and it is possible that the interaction effects found in this study are aggregations of these more specific patterns within condition or drug-specific subpopulations. Future research should explore causes of these effects, since this may inform strategies to improve care for patients on polypharmacy, including where they may be affected by health inequalities. Such strategies should be informed by real time analytics driven by routinely collected electronic health record data, to help focus intervention towards those with greatest need.

In this study we did not restrict attention to problematic polypharmacy, but rather set out to quantify the overall prevalence of polypharmacy and to understand variation by demographics and frailty. Future research should extend this analysis to examine problematic polypharmacy specifically, and quantify costs associated with problematic polypharmacy accounting for patient harm resulting from adverse events. This should include a focus on commonly co-prescribed medications and their interactions, especially any potentially inappropriate combinations as prescribed in practice in the population. Furthermore, benefits of medication should be incorporated into health economic analyses to provide the full picture.

Our method for counting distinct medications prescribed to each patient using primary care records may be used both in future research on polypharmacy, and in practice by regional or national health systems to measure and analyse polypharmacy with a view to improving quality of care.