Identifying frequent drug combinations associated with delirium in older adults: Application of association rules method to a case‐time‐control design

Older adults are at an increased risk of delirium because of age, polypharmacy, multiple comorbidities, frailty, and acute illness. Although medication‐induced delirium in older adults is well understood, limited population‐level evidence is available, particularly on combinations of medications associated with delirium in older adults.


| INTRODUCTION
The association rules (AR) methodology is a data-mining algorithm that extracts from a big dataset frequent and statistically interesting item sets above a chosen frequency threshold. [1][2][3] The AR method has been successfully used in pharmacoepidemiology and post-marketing surveillance to investigate frequent medication and medication combinations contributing to adverse drug events (ADEs). 1,4,5 The AR method has also been used to illustrate the complex interactions of multimorbidity, depict common comorbidity patterns, and assess medication use complexity in community-dwelling older adults. 4 The utility of the AR method has also been successfully extended to the field of bioinformatics to identify factors that control gene transcription. 6 We previously have demonstrated the AR method's utility to investigate medication combinations associated with fracture and acute kidney injury in older adults. 7,8 Our previous analyses were restricted to examining transient medication exposures during the time at risk of an acute event in line with the recommendation for implementing a case-crossover design. In this study, we used a case-time control design to adjust for time trends of medicine exposures.
This study chose to apply the AR method to a case-time-control design as it mitigates confounding from unknown time-invariant confounders. We followed all the recommendations to apply a case-time-control design to our analyses. 28,29 The key assumptions met include that the occurrence of the event must be acute, and the exposure may vary over time.
The overarching aim of this case-time-control study was to apply the AR method to identify frequent medication combinations contributing to the risk of delirium in older adults aged 65 years and older.

| METHODOLOGY
We obtained ethical approval from the Human Ethics Research Committee, University of Bath (approval number EIRA1-5353).

| Data source
The New Zealand (NZ) Ministry of Health holds national collections of the community pharmacy dispensing, hospital discharge details and mortality data details. Pharms is the national collections of all prescription claims made by community pharmacists. It contains prescriptions of medicines funded by the Pharmaceutical Management Agency (PHARMAC). PHARMAC is the New Zealand government agency that decides which pharmaceuticals to fund in New Zealand publicly and provides funded access to pharmaceuticals for all New Zealanders.
The National Minimum Data Set (NMDS) is a national collection of public and private hospital discharge information, including coded clinical data for inpatients and day patients. We have provided a detailed description of both the datasets previously. 7,8 We used unique encrypted National Health Index (NHI) identifiers to cross-match medication exposure data with hospital events data from NMDS.
We were interested in exploring the use of all PHARMAC-funded drugs and combinations of these at the population level, including mainly antimicrobials, antihistamines, antihypertensive drugs, anticholinergic drugs, antidepressant drugs, antipsychotics, benzodiazepines, diuretics, opioids, and nonsteroidal anti-inflammatory medications. 2

| Case-time-control cohort
We created a case-time-control cohort for medication exposures, with The case period is the day before the index date. The control period is the l-day period 36 days before the index date. We chose two 5-week washout periods based on our previous study to avoid carrying over the effects of any medication exposures around the control period and to exclude prescriptions that are dated and unlikely to have effects within the observation period. 8 We calculated the duration of each prescription by dividing the total dose supplied by the daily dose. We determined whether an individual had nonintermittent exposure to the drugs of interests within the case and control periods with the prescription dates.

| AR methodology and statistical analysis
We applied AR methodology in this study to identify exposures to drugs and drug combinations associated with an increased risk of delirium. In this study, drugs and drug combinations that individuals were exposed to with a frequency of at least one in every 200 individuals on the day before the event (i.e., within the case period) are the frequent itemsets. The interestingness statistics are the increased odds of delirium onset for each frequent drug combination due to the exposure.
We expressed the increased odds of delirium onset due to exposures as matched odds-ratio (MOR). We identified individuals with exposures to medication combinations in the case and the control periods from the case-time-control cohort. We counted the number of individuals, with delirium who were exposed to drugs of interest We used heat maps to display the strength of the association between medication exposures and delirium. In the heat map, each row represents an exposure combination, and each column represents a medication that appears at least once in the set of exposure combinations. If a medication appears in a particular exposure combination, the corresponding grid is coloured. We then mapped the MOR calculations onto the heat map. Each row (i.e., each exposure combination) is proportional to the log(MOR) of delirium associated with this combination. The grid bordered with a blue colour in the heat map represents log(MOR) > 0 with a confidence level of 95%.

| Sensitivity analyses
We conducted sensitivity analyses to account for the carrying-over effects of drug exposures. We selected two other periods lengths, 3 and 7 days, for the case and control periods instead of 1 day. Casetime-control cohorts were re-created with those period lengths, then, the AR as mentioned above analyses were repeated.
A longer time window accounted for the possibility of carryover effect contributed by discontinuation closer to the observation period.
When observation periods are longer than a day, we defined nonintermittent exposure to medication if prescribed for more than 80% of the days within the study period. The 80% cut-off is considered a standard measure for medication adherence in pharmacoepidemiological studies (i.e., the proportion of days covered ≥0.8).
The pharmaceutical collections (PHARMS) and NMDS data were made available as annual, CSV-formatted datasets. The filtering mentioned above and cohort-construction procedures were performed using a computer program written in R (3.4.2, R Core Team, 2016).

| RESULTS
We identified 28 503 individuals (mean age 84.1 years) from 2005 to 2015 with a recorded diagnosis of delirium. Of these, 24 814 had at least one prescription involving PHARMAC-funded drugs within the case period, the day before the delirium event. The number of individuals exposed to each of the 41 exposures mentioned above is shown in Table 1. For most combinations, the distribution of ages was slightly skewed towards the higher age group.  Table S2).
The sensitivity analyses were repeated with (3-day) and weekly    evidence from a prospective cohort study conducted in older hospital inpatients. 32 This study found that benzodiazepines accounted for one-third of delirium cases in patients with normal cognitive impairment.
The AR method also revealed loop diuretics increase the risk of delirium. The use of diuretics can be associated with fluid and electrolyte imbalance, including hyponatremia resulting in the onset of confusion and delirium. 20 In a subgroup analysis (cohort size approximately 91 000) of a matched cohort study using electronic health records, calcium channel blockers were associated with less delirium than diuretics (OR 0.84 (079-0.90)). 33 The finding that fentanyl increases delirium risk in older adults is biologically plausible and congruent with evidence that links opioid use with delirium. 34,35 The one postulated mechanism is that opioids induced delirium secondary to their central anticholinergic effects 36,37 and directly acting on the opioid receptors. Pre-clinical studies conducted in rats have shown dose-dependent toxicity associated with fentanyl possibly mediated via binding to the brain's muscarinic receptors. 38 A case report also highlighted delirium risk in a 71-year-old elderly patient following fentanyl exposure. 39 Our study found a combination of furosemide, omeprazole and lorazepam increases the risk of delirium. There is a lack of welldesigned studies addressing medication class combination associated with delirium. This study provides new insights into medication combinations that are linked with delirium. However, it should be noted from case reports that individually omeprazole, 40 furosemide and lorazepam 41 may increase the risk of delirium.
The finding that quetiapine is associated with delirium is interesting as antipsychotic medications are recommended to treat delirium. 42 Antipsychotics are known to be associated with delirium, potentially mediated by their anticholinergic properties. 43 Case reports have shown quetiapine and its metabolite norquetiapine have strong anticholinergic effects and may induce delirium in older adults. 44 Case reports have also highlighted delirium associated with olanzapine secondary to its anticholinergic effects. 26 The sensitivity analyses repeated with (3-day) and weekly (7- However, the effects of time-varying confounders that change rapidly are difficult to adjust, which is a major limitation of this study.
For example, acute health status changes, such as dehydration, may also increase delirium risk. 45 We mitigated time-varying confounding by studying only the short, 72-day time period before the index date, but residual confounding remains a concern. The other limitation is protopathic bias; haloperidol and lorazepam are often recommended to treat delirium. [46][47][48][49] Since we used pharmacy claims data, exposure misclassification due to lack of information on medication consumption, self-medication, and the use of over-the-counter drugs such as NSAIDs linked to delirium may have biased the findings. 50 The NZ pharmacy claims data includes medications funded by PHARMAC and may differ from other countries' drug formularies, and this limitation may affect the generalisability of the study findings.
We met the assumptions that the outcome of interest must be an acute event and not change over time; delirium is an acute event and is unlikely to change over the short study period employed in our study. However, we should be conscious of other important assumptions and limitations of the case-time-control design, including bias created by selecting the case and control windows, selecting a control group, and inadequate adjustment of timevarying confounders. 29,51 We also acknowledge that persistent user bias cannot be eliminated by using case cross over designs studying chronic exposures as it may bias the findings (odds ratio) upwards. 52 We applied a 5-week washout period universally for all exposures, and the differences in half-lives of individual medicines were not considered.

| CONCLUSION
The association rule method applied to a case-time-control design is a novel approach to identifying drug combinations contributing to delirium with adjustment for any temporal trends in exposures. The study provides new insight into the combination of medicines linked to delirium.