No Association Between Pharmacogenomics Variants and Hospital and Emergency Department Utilization: A Mayo Clinic Biobank Retrospective Study

Introduction

Healthcare organizations utilize risk stratification as an important tool for managing population health.^2,3 In the ambulatory practice, clinicians determine high-risk patients by their age⁴ as a primary predictor of risk for hospitalization. Risk prediction instruments utilize other risk factors for hospitalization including comorbid health conditions,^5–7 and self-reported physical health.^4,8 Other important risk factors for health outcomes include socioeconomic status⁹ and previous hospital utilization.⁵ Following the determination of high-risk status, healthcare organizations might provide different levels of care to best suit the risk stratification of the patient.¹⁰ While the traditional methods of risk stratification for hospital use are effective,¹¹ further refinement of risk would improve our ability to better identify patients at high risk. Pharmacogenomics is one potential clinical factor that may help predict adverse health outcomes given the association between medication and adverse events like emergency department visits.¹² To date, personalized medicine using genomic information has not been used in risk-stratification tools for health outcomes including hospitalization.

While pharmacogenomics has not been specifically evaluated for healthcare utilization, the use or misuse of medications can certainly place patients at risk for hospitalization or ED use. Polypharmacy may place patients at risk for adverse health outcomes like 30-day hospital readmission¹³ or hospitalization in patients receiving home care.¹⁴ Patients experience an increased risk of ED visits when they take pain medications, diabetic medications, or anticoagulants.¹⁵ The American Geriatrics Society published guidelines on medications that are potentially hazardous for older adults.¹⁶ Providers recognize common inappropriate medications like anticholinergic medications, opioid pain medications, sedative/hypnotic medications among other medications.¹⁶

The metabolism of medications could potentially affect efficacy and toxicity¹⁷ and may impact the majority of patients. A recent study tested only five pharmacogenomics (PGx) and yet found that 99% of patients had one or more actionable results.¹⁸ In our previous study of 1013 patients with pharmacogenomics for CYP2D6, we found 8% of patients possessed a poor metabolizer phenotype and 8% had ultra-rapid metabolizer phenotypes.¹⁹ In further work in this same population, we showed that the extremes of metabolism phenotype (poor or ultra-rapid) were associated with a higher risk of hospitalization.²⁰ However, this has not been shown in all studies that have examined pharmacogenomics phenotypes and hospitalization. In a study of 729 patients with pre-emptive VKORC1 and CYP2C9 PGx genes, pharmacogenomics-guided warfarin dosing did not lead to a reduction in hospitalization or mortality.²¹ In addition, the identification of new alleles in pharmacogenes plus improved variant functional characterization has led to new variant classifications.^22,23 In this study, we undertook a larger examination of the association of pharmacogenomics phenotype agnostic of specific medication usage and the utilization outcomes of hospitalization and ED visits using the most recent variant classifications. Further, we explored the association between pharmacogenomics and disease burden because as many as 50% of patients react inadequately to prescribed drugs²⁴ which can lead to poorly treated health conditions.

Methods and Materials

Study Subjects

This is a retrospective cohort study. All participants in this study were enrolled in the RIGHT expansion study as described previously by Bielinski et al.¹ Briefly, adult subjects were selected from the Mayo Clinic Biobank (MCB)²⁵ for use of the Mayo Clinic for their healthcare and offered the opportunity to receive pharmacogenomic testing plus deposition of key results in the Mayo Clinic electronic health record (EHR). A total of 18,199 adult subjects were invited and 10,085 enrolled. Subjects who had subsequently withdrawn consent from the MCB or the RIGHT 10K study were excluded, leaving 10,078 in these analyses. The majority of RIGHT 10K participants lived within the Rochester Epidemiology Project (REP) catchment area where we can access comprehensive electronic health records (EHRs) information including hospitalization and ED visit information.²⁶ The Mayo Clinic IRB and the Olmsted Medical Center IRB reviewed and approved the study protocol. The study was conducted in accordance with the Declaration of Helsinki. The study was also reviewed and approved by the Mayo Clinic Biobank access committee and the RIGHT access committee.

Outcomes

We searched the EHR systems of Mayo Clinic and Olmsted Medical Center to ascertain hospitalization and/or ED visits. Follow-up time began on the date of consent to the Mayo Clinic Biobank (April 2009 through April 2017). Follow-up time ended when participants exited the study which was either the date when PGx gene information was available in participants’ EHR or the date of death date, whichever came first. The median time for the patients to exit the study was 7.3 years to study exit. Index date was defined as the first date (or the first date if multiple outcomes occurred) associated with ICD 9 or 10 codes for each outcome. Hospitalization included a stay that was overnight and excluded outpatient hospital services (ie, colonoscopy and outpatient surgery). We used the REP data linkage system to identify participants with each outcome. The medical cause for readmission was identified using ICD9 and ICD 10 codes and was mapped to Phecodes.^27,28

Predictors

We obtained sociodemographic characteristics from the EHR which included age, sex, and race (white race versus non-white race) and the MCB questionnaire for educational level and self-rated perceived general health. Educational level was reported as high school or less, some college or 4-year college degree, advanced degree, or missing. We reported the classification of patients’ self-reported perceived general health as excellent/very good, good, fair/poor, or missing. We obtained chronic condition information at the time of the biobank consent using the Department of Health and Human Services guidelines for determining chronic conditions²⁹ which utilized previously methodologies and added them to define disease burden.³⁰ The conditions were derived from ICD 9 or 10 billing codes for all participants.³¹ The team quantified socioeconomic status using the HOUSES index which uses housing characteristics.³² We placed the HOUSES index scores into quartiles for comparison consistent with our previous work with higher quartiles with higher socioeconomic status.⁹

For pharmacogenomics phenotypes, we utilized the records of those patients with PGx data within the biobank. The PGx phenotype was obtained from PGx sequencing data from the RIGHT protocol. These PGx genes were selected based upon the best clinical practice advisories and decision support for clinicians. We reported all phenotypes in Supplementary Table 3; however, because of the small numbers of phenotypes and outcomes, we categorized phenotypes into normal, abnormally fast, or abnormally slow. For DPYD, we reported as normal, intermediate or poor. For HLA, we reported present or absent.

Statistical Analyses

Socio-demographic data were summarized as counts and percentages or medians and 25–75%tile as appropriate, stratified by status of each outcome. Differences in risk of hospitalization from PGx phenotypes were assessed by modeling time from biobank consent to hospitalization predicted by PGx phenotypes using Cox proportional hazards models. We analyzed the primary endpoint of hospitalization using Likelihood ratio tests based on comparing Cox models with and without the individual PGx phenotype predictors. We also tested whether PGx phenotype is associated with disease burden and/or sex, known risk factors for risk of hospitalization outcomes, by using Kruskal–Wallis test and Chi-square test, respectively. In a similar fashion, we used Cox proportional hazard model for ED visit outcome.

Results

Of 10,078 patients, the average age of the patients at the MCB enrollment was 66 years with (IQR: 54 to 74; Table 1). Consistent with prior literature, patients who were hospitalized compared to non-hospitalized, were older (median 67 years versus 65 in non-hospitalized, p value <0.001) and were more likely to have high school or less education (22% versus 13%, p value <0.001). Female gender was also less likely to be hospitalized compared to males (OR 0.76 (95% CI: 0.68,0.84)). Patients of white race were also less likely to be hospitalized compared to non-white race (OR 0.73 (95% CI 0.56,0.96)). The rate of hospitalization and ED visit at 5 years were 9.3% and 5.4%, respectively. The hospitalized patients were more likely to report poor/fair health (15%) compared to non-hospitalized patients (5%) (p value<0.001). Those hospitalized tended to have lower socioeconomic status measure by HOUSES (18% vs 12% for the lowest quartile in the local population, p value <0.001), which is similar to the trend associated with education level. (Table 2) When looking at ED visit as the primary outcome, we found similar associations between ED visit and age, sex, race, educational level, self-rated health and socioeconomic status (Table 3).

Table 1 Characteristics of Study Subjects Included in the Study

Table 2 Association Between Subject Characteristics and Risk of Hospitalization Status During Follow-Up Period

Table 3 Association Between Subject Characteristics and Risk of Emergency Department (ED) Status During Follow-Up Period

We did not find any association between the phenotype of 10 PGx genes and hospitalization. We have reported all of the phenotypes for the 10 pharmacogenomics genes in Supplementary Table 1. In comparing the phenotypes of frequencies of CYP 2D6 to other world populations, they appear similar.³³ There was very little variation in phenotype between those with hospitalization and those without hospitalization (Table 4). Specifically, in CYP 2D6 those with a fast phenotype had 2.6% hospitalization compared with 2.0% in those without hospitalization (p 0.40). For our second outcome, we found that there was no association between ED visits and pharmacogenomic phenotype (Table 5). Of the pharmacogenomic phenotypes, dihydropyrimidine dehydrogenase (DPYD) had a trending p value of p=0.06. We report no difference in pharmacogenomics phenotype by sex or disease burden (Supplementary Table 2), which justifies the rationale for not adjusting for these variables when testing the association between PGx phenotype and the outcomes. For the primary diagnosis for admission, we found osteoarthritis and major depression as the top two reasons for admission (Supplementary Table 1).

Table 4 Univariate Association Between 10 Selected Pharmacogenomic Phenotypes and Risk of Hospitalization. Note 48 Subjects (8 with Hospitalization; 40 without Hospitalization) Had No Information on Pharmacogenomics Phenotypes

Table 5 Association Between 10 Selected Pharmacogenomic Phenotypes and Risk of Emergency Department (ED) Visits. Note 48 Subjects (3 with ED Visits; 45 without ED Visits) Had No Information on Pharmacogenomics Phenotypes

Discussion

In this study of 10,078 patients, we did not find a relationship between PGx phenotype and hospitalization and ED visits. In particular, we did not find a relationship between the common PGx genes of CYP 2D6, CYP 2C19 and CYP 3A4 and hospitalization and ED visits. This is a novel finding as this type of study has not been performed previously and provides some initial evidence on the utility of PGx phenotype to predict hospitalization or ED visit. These results reinforce that traditional risk factors for hospital readmission like age, sex, race, comorbid health burden,³⁴ previous utilization,^34,35 and education³⁶ are stronger predictors of hospitalization. Our previous work also shows the importance of these traditional risk factors.^5,11,37 PGx phenotype as currently defined does not help predict hospitalization or ED visit.

In our previous study, our data indicated that 8% of the cohort had ultra-rapid phenotypes for CYP 2D6 and that this phenotype was associated with hospitalization outcomes.²⁰ In contrast, in the current study, we found only 2.1% of patients had ultra-rapid phenotype. The differences most likely reflect re-evaluation of the activity of the *2A allele that occurred in the interim period by the laboratory that conducted the phenotyping assays, suggesting that our previously reported finding was a chance finding.³⁸ The continual improvement of pharmacogenomic assays is critically important as we work to improve the understanding of pharmacogenes, especially CYP2D6. The metabolism phenotype for CYP 2D6 has numerous clinical implications including antipsychotic medications,³⁹ antidepressants⁴⁰ and pain medication like oxycodone or codeine.⁴¹ In fact, CYP 2D6 is known to affect the metabolism of approximately 25% of the known drugs.⁴²

Clinical medicine is just starting to understand personalized medicine and PGx. There have been some common uses of PGx for individual medications and PGx phenotypes. Clopidogrel and CYP 2C19 is one such common example as clopidogrel is a prodrug and is commonly used as an anti-platelet medication to prevent stent re-thrombosis.⁴³ In a meta-analysis of studies of clopidogrel and CYP 2C19, there were worse health outcomes with reduced function phenotypes with hazard ratios of 1.55; 95% CI, 1.11–2.17; P = 0.01 for patients with one reduced function allele compared to normal types for cardiovascular death, stroke or myocardial infarction.⁴⁴ In a study of patients starting medications for mood disorders, pre-emptive PGx testing resulted in lower costs and utilization.⁴⁵ In another observational study, investigators found lower hospital length of stay in patients receiving pharmacogenomics testing compared to no testing in depressed patients.⁴⁶

The strengths of this study included a comprehensive list of outcomes (hospitalization and ED visit) from the REP data system and catchment area.⁴⁷ It included the pharmacogenomics information which was unique and integrated into the medical record.⁴⁸ The limitations included the potential for missing outcomes or predictors if the patient sought care outside the REP catchment area; however, this is minimized because the largest health systems are included for the outcomes. In this cohort study, there may be unrecognized confounders that can influence hospitalization or ED visits. The study does not report drug-gene pairs, and we cannot assume that hospitalization or ED visits were associated with a complication of medication or of a problem that PGx can solve. We also accept that the population of the cohort is largely white¹ which may generalize to the upper Midwest of the United States but not to other regions of the world.⁴⁹ Additionally, we acknowledge that we may lack statistical power as some PGx phenotypes had relatively small sample sizes in abnormal categories. However, the observed effect size was in general small in the majority of the PGx phenotypes.

Conclusion

We found no association between pharmacogenomics phenotype and hospitalization or ED visit in patients who had pharmacogenomics testing. This lack of association adds to our knowledge about the effect of individualized medicine and risk stratification for hospitalization and ED visit. Traditional risk factors for hospitalization including age, comorbid health concerns and previous hospitalization^4,5 will still be standard methods to help predict future hospitalization. Pharmacogenomics will continue to play instrumental roles in preventing adverse drug reactions in high-risk patient populations.¹⁷