Older adults living with dementia in community settings have approximately three times the odds of being prescribed five or more medications (polypharmacy) or ten or more medications (hyperpolypharmacy) compared to those without dementia [1]. The risk of adverse events due to drug-drug and drug-disease interactions [2, 3] may be amplified in older adults living with dementia due to a high number of comorbidities [4], increased frailty [5], and pharmaceutical management of symptoms. Cognitive impairment can make it difficult to manage multiple medications with varying prescribing regimens and challenges articulating adverse effects [6, 7]. Inappropriate prescribing can also exacerbate cognitive decline [8]. Understanding medication prescribing and use in this population is a research priority [9] and could inform strategies to reduce polypharmacy and support deprescribing initiatives.

Polypharmacy and hyperpolypharmacy in older adults have been extensively explored [10–19]. The use of cross-sectional studies and binary exposures (e.g., yes/no having polypharmacy) have been highlighted as limitations in current studies examining medication-related adverse events [20]. These approaches may not capture complexity contributing to medication related adverse events adequately, and do not account for changes in prescribing patterns over time or disease progression. There is an opportunity to build on current methods using unsupervised machine learning to characterize prescribing.

Network analysis is a graphical method, frequently applied in social science and ecology [21–23], where a network is a set of objects, called nodes, with connections among them, called edges or links [24]. This method can be applied to medication data, where medications are considered nodes and medications an individual uses concurrently are linked through edges. Network analysis graphs can be used to map medications, providing a visualization of prescribing patterns. Networks can be summarized quantitatively using metrics associated with the number of nodes, size of edges among nodes, and measures of node importance (i.e., centrality). These measures have been used to compare complexity of different ecological systems [21, 25], the strength of connections within groups (e.g., genuses) [26], and highlight important “players” within a network [27]. These measures can all be applied in the context of medication networks to compare the medications prescribed to different subgroups (for example, of different patient and prescriber characteristics), to examine the strength of co-prescribing patterns and, possibly, to identify key medications contributing to polypharmacy. A small number of studies have applied network analysis in a pharmacoepidemiology context [28–30]; however, analysis of medication networks has not yet been applied to the context of community-dwelling older adults with dementia.

The goal of this paper was to conduct a network analysis of dispensed medications of community-dwelling people living with dementia across two time points to illustrate this method and its advantages in understanding population-level prescribing.

We identified 99,064 individuals living with incident dementia, and 15,222 were alive and not living in a nursing home after five years (n=35546 [36%] dying during follow-up and n = 48,278 [49%] being admitted to long term care during follow-up). At index, the mean age was 83.1 years and 57.7% of the cohort were female. More than half (53.6%) had over five chronic conditions, including hypertension (82.5%), history of cancer (56.8%), and coronary heart disease (37.2%). (Supplementary Table 3)

It is well-known that patterns of medication use can be complex in individuals living with dementia [1, 6, 14, 43–45]. This study offered an introduction to the fundamental aspects of network analysis by examining dispensed medications in a cohort of individuals with incident dementia who were examined at two cross-sectional time points. We used network graphs and associated metrics to explore how this method can be used to characterise and compare prescribing when diagnosis is ascertained in the health administrative data and five years after. The most frequently dispensed medications, primarily those used to treat and manage cardiovascular disease, remained at similar proportions at index and at five years. There were changes in some CNS-active medications through time. Medication networks after five years were denser, which could suggest medication use became more complex through dementia progression, although our findings also suggest selection in who remains living in the community over time. Node-level measures of centrality highlighted frequently prescribed medication subclasses, such as statins, proton pump inhibitors, and beta blockers, as well as CNS-active medication like SSRIs and other antidepressants.

One of the main strengths of network analysis for understanding population prescribing patterns are the visual graphs. In a network graph, key frequencies and concurrent dispensing were identifiable. For example, graphs showed high usage of cardiovascular medications. This medication class is often overlooked in research considering potentially inappropriate prescribing in older adults because of the strong safety and tolerance profile of cardiovascular medications [46]. Despite lower risk profiles relative to CNS-active medications, a narrative review found up to 50% of medication-related harms in older adults may be attributable to cardiovascular medications that is often under-recognized in clinical practice [46]. The appropriateness of cardiovascular medication may change over time in older adults with dementia, and a review highlighted the need for additional research [47]. The role of cardiovascular medications needs to be considered in drug-drug interactions, drug-disease interactions, and polypharmacy at a population-level and in individual medication reviews, because of their high prevalence of use in community-dwelling older adults living with dementia [48, 49]. Thus, the graphical aspect of network analyses can complement traditional frequentist statistical approaches as it demonstrates the complexity of prescribing. For example, the information included in the graph representing the top 25 medications reflects a 25 by 25 matrix of the number of co-prescriptions between each medication subclass, a frequency list for 25 medication subclasses, and the ATC classification for each of the 25 medications.

Another important benefit of plotting the medication subclasses as nodes was being able to readily compare two graphs across populations – in our example, medication use at two different timepoints. Many of the CNS-active medication subclasses were more frequently prescribed at five years, which was seen in the increased node size, consistent with CNS-active medication use increasing through disease progression in people living with dementia [6, 50]. Previous studies have also used networks to highlight comparisons across groups. For example, networks stratified by age and sex found differences in patterns between females and males aged 30 to 60 years, but not in other age groups [29]. Networks for females were more interconnected which authors suggested may be due to a larger number of pathologies and higher pathophysiological complexity [29].

When there are many nodes with multiple connections there is limited utility to the visualization produced when mapping all nodes and edges in a single graph. Network graphs are often used to plot more sparsely connected nodes, such as social networks or in ecology [25, 26]. Restricting the networks, either by frequency to the most common nodes or by ATC category, made the visualizations understandable. In our study, medications were grouped into subclasses because networks at the drug name (i.e., generic medication) level would have been even more dense and less interpretable. Many of the documented (e.g., American Geriatric Society Beers criteria [42]) potentially inappropriate prescribing patterns occur at the drug therapy level, so that when investigating drug-drug interactions, producing networks at this more granular level may be necessary.

Networks are more than a graphical representation of relationships in data. Attributes of the networks – such as metrics about nodes, edges, and centrality – can provide insights when applied to medication data and used to make comparisons across patient subgroups, prescribers, health care settings (i.e., nursing homes) and regions. In our example, metrics derived from the network changed between time periods, suggesting more complex prescribing to those still surviving in the community at five years. Edge metrics in medications networks may be of particular interest when the co-prescribed pairs themselves are important, as in the example above when investigating medication pairs constituting potentially inappropriate prescribing for older adults or those living with dementia [42, 51, 52]. In very dense and highly clustered medication networks in older adults in Italy [28], proton pump inhibitors were more frequently co-prescribed than expected with other proton pump inhibitors [28], especially since these prescriptions often do not have a clear clinical indication in this population [53].

There were limitations to the practical utility of metrics in our example, but which may be more useful to other research studies applying this method. For our population, there was relatively little difference in the dispensation of medication subclasses between index and 5 years, and thus limited changes in the associated metrics. However, differences in those who remained in the cohort (i.e., alive and not living in a nursing home after five years) likely contribute to this. Several other studies have examined the change in medication use over time in people with dementia, and reported increased, decreased, or no change in proportion of medication subclasses being dispensed. For example, Gnjidic et al examined the impact of dementia diagnosis on potentially inappropriate prescribing, finding an 11% increase in number of medications one year following diagnosis – leading to a 1.2 higher odds of a potentially inappropriate medication [44]. Other studies found minimal changes or no changes in medication subclass use [48], increase in drug-drug interactions or potentially inappropriate prescribing [48] or association between polypharmacy or hyperpolypharmacy with a trajectory of cognitive decline [54] after a dementia diagnosis. This purpose of this study was to examine the usefulness of networks when describing population-level dispensing in older adults with dementia, and thus we did not delve deeply into the clinical appropriateness of medication pairs.

High node-level centrality highlighted medications that were frequently dispensed. Although this validates these network measures in their ability to select important nodes, it suggests complicated calculations of centrality scores may be less useful for densely connected networks if the same information can be obtained from frequency tables. Additionally, in highly connected networks, nodes often have highly associated betweenness, degree, and eigenvector centrality scores [40], suggesting researchers can choose one centrality score that is appropriate for the research question and interpretable for their data. In our study, medication subclasses with high degree centrality are frequently concurrently dispensed with other medications and may be important to highlight in medication reviews. Although not seen in our example, in networks with multiple distinct clusters (also called communities), medications with high betweenness centrality could indicate areas for clinical intervention to reduce polypharmacy. Similar “bridges” have been identified in studies of infectious disease transmission using network analysis [25] where intervention could stop or limit outbreaks. The clusters which could show groups of medications more likely to be prescribed together (i.e., more than medication pairs) were not clear in the commonly prescribed medications, which is consistent with densely connected networks.

Network analysis can be used to reveal both well-known and under-recognized prescribing patterns occurring at a population-level, demonstrate changes over time and/or between subgroups, and might suggest areas to optimize prescribing across populations. This study showed how this method can compare multiple aspects of prescribing simultaneously, such as the percentage of cohort being prescribed a medication, the strength of co-prescribing, and the overall complexity of prescribing. There are extensions of the method, such as community detection, not explored in this paper, that may be useful in future applications. Continued exploration of techniques novel to pharmacoepidemiology are important for the advancement of our understanding of prescribing in those living with dementia in the community, and in turn, improving medication use in this population.