https://orcid.org/0000-0001-9464-9123
Figures
Abstract
Background
Data on drug safety in multimorbid older-adults are limited, as clinical trials often apply upper age limits and focus on individual drugs or specific combinations. We aimed to explore high-risk drug patterns in community-dwelling older-adults, and their associations with long-term mortality.
Methods
We included 1,048 participants from a longitudinal population-based cohort, all taking at least one medication. Participants were examined in 1999–2007 and followed for mortality through March 2022. Individuals with similar profiles of high-risk drugs, identified using Beers criteria as potentially inappropriate for most older adults or requiring caution, were grouped using agglomerative hierarchical clustering. Cox and competing-risk regressions were used to examine the associations of the high-risk drug patterns with all-cause and non-cancer mortality.
Results
The most prevalent morbidities among participants (mean age 73.3 ± 7.3 years, 55.9% women) were hypertension (55.3%) and cardiovascular diseases (45.5%), and 77.7% took at least one high-risk drug. Five distinct patterns were identified: ‘None’ cluster (no dominant high-risk drug); ‘Calcium channel blockers’ (CCBs) cluster, with high nonsteroidal anti-inflammatory drug (NSAID) prevalence; ‘Renin-angiotensin-aldosterone system (RAAS) inhibitors’ cluster, with a high concomitant use of sulfonylureas compared to other clusters; ’Diuretics’ cluster, with a relatively high prevalence of antithrombotics and proton pump inhibitors; and ’Benzodiazepines’ cluster, with a relatively high antidepressant prevalence. Clusters differed by age, sex, ethnicity, and health characteristics. In multivariable analysis, the ‘Diuretics’ cluster was associated with increased all-cause (HR = 1.33, 95%CI: 1.03–1.72) and non-cancer (HR = 1.41, 95%CI: 1.03–1.93) mortality compared to the ‘None’ cluster. The ‘CCBs’ cluster was associated with a greater risk for non-cancer mortality. Several drug combinations were identified as potential contributors to the increased risk observed in these clusters, including the concomitant use of NSAIDs and antihypertensives and a possible CCB-diuretic prescribing cascade.
Citation: Orenstein L, Chetrit A, Fluss R, Laufer K, Szklo M, Dankner R (2025) Associations of high-risk drug patterns with mortality among community-dwelling older adults: A 23-year prospective cohort study. PLoS One 20(9): e0332210. https://doi.org/10.1371/journal.pone.0332210
Editor: Sepiso K. Masenga, Livingstone Center for Prevention and Translational Science, ZAMBIA
Received: March 31, 2025; Accepted: August 27, 2025; Published: September 11, 2025
Copyright: © 2025 Orenstein et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Data cannot be shared publicly due to privacy, ethical and legal restrictions, as the dataset contains potentially identifying and sensitive patient information. Researchers interested in accessing the data may submit an application via the TIMNA platform (https://govextra.gov.il/ministry-of-health/big-data-research/registration/) or contact the Ministry of Health institutional TIMNA staff at timna@moh.gov.il. Alternatively, researchers may contact the corresponding author for guidance on how to apply for access, which is subject to institutional and regulatory approvals.
Funding: This work was supported in part by Tel-Aviv University, Tel-Aviv, Israel (through the Albert and Alba Cuenca Institute for Therapeutical-Approaches to Age Related Diseases; https://en-med.tau.ac.il/About-Sackler-Faculty-of-Medicine), and by the Israel National Institute for Health Policy Research (grant number 180/2018; https://www.israelhpr.org.il/en/). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Older adults often present with multiple chronic conditions [1]. As most clinical guidelines are single-disease based, multimorbidity leads to long-term treatment with multiple drugs [2]. However, high-quality data on drug safety and efficacy for this growing population are limited, largely due to their frequent exclusion from clinical trials [3–5]. Thus, guidelines often disregard the potentially different pharmacokinetics and pharmacodynamics due to age-related alterations in body composition and metabolism, the greater heterogeneity in terms of physiological reserves, and the complexities of treatment within the context of multimorbidity and polypharmacy [6,7]. Older adults are consequently more susceptible to adverse drug reactions (ADRs) than younger individuals [6,8]. Recognizing this public health concern, various tools have been developed to identify potentially inappropriate medications (PIMs) in older persons aged 65 or above, i.e., drugs that are ineffective or pose unnecessary risks [9]. The Beers criteria, a widely accepted tool, that is periodically updated with the American Geriatrics Society [10–13], classify drugs to avoid or be used with caution in older adults. Some of these drugs are associated with high rates of hospitalization (e.g., antipsychotics [14–16], antidepressants [14,16], nonsteroidal anti-inflammatory drugs (NSAIDs) [14,17], diuretics [14,18], benzodiazepines (BZDs) [14,16,19], opioids [15,20] and anticholinergics [21,22]). Additional drugs, e.g., calcium channel blockers (CCBs), renin-angiotensin-aldosterone system (RAAS) inhibitors and antithrombotics, were identified in the literature as “high risk” owing to their consistent association with increased adverse events in older adults [6,14,18,23].
The impact of drug regimens on mortality has been less studied. Some drugs are consistently associated with increased mortality risk (e.g., anticholinergics [21,24] and opioids [15,25]), whereas results vary for other drugs, such as antipsychotics [15] and BZDs [24,26]. Conversely, drugs such as CCBs [27] and RAAS inhibitors [28] are associated with decreased mortality.
Older multimorbid patients often use various high-risk drug combinations, yet limited literature exists on these patterns [20,29], and their associations with mortality [23]. In addition, widely used administrative databases do not capture over-the-counter (OTC) drugs, such as NSAIDs [15], which can interact with prescribed drugs and impact safety and adherence [30,31]. Moreover, self-medication has been associated with increased risk of drug interactions [32].
As the population ages, high-quality data are crucial for developing evidence-based medication guidelines for older adults. This study aimed to examine patterns of high-risk drug use, considering both prescription and non-prescription regimens, in a population-based cohort of community-dwelling older adults; and to further investigate their associations with all-cause and non-cancer mortality. Taking a patient-centered approach, we evaluated the combined use of high-risk medications rather than focusing on individual drugs, to better reflect real-world use and identify profiles of high-risk patients who could benefit from interventions targeting their medication use.
Methods
Study population
We used data from Phase 3 of the Israel Study of Glucose Intolerance, Obesity and Hypertension (GOH Study), a longitudinal study that began in 1969, with a sample of 5,711 Jewish participants, stratified by sex, age and ethnic origin. The current study is based on data collected during Phase 3 (January 1999 through November 2007), the only wave in which detailed medication information was collected. The final study cohort included 1,048 participants who were approached for the third follow-up, excluding those not taking any medications (n = 162). Face-to-face interviews covered sociodemographic characteristics, lifestyle habits and medical history; anthropometric measurements and blood samples for biochemical and genetic biomarkers were taken. The study was approved by the local IRB of the Chaim Sheba Medical Center (Approval # SMC-8394-21). All participants signed a written informed consent form upon recruitment to the original GOH Study. Since only anonymous data was used in the current study and patient privacy was not violated, obtaining informed consent for the current study was waived by the IRB (decision made on August 5, 2021).
High-risk medications
During the baseline interview, participants presented the packaging of all medications they regularly take. Drug information (name, indication, dosage, duration) was cross-checked against a summary letter from their physician, then recorded and coded using the Anatomical Therapeutic Chemical (ATC) classification. ATC codes were assigned based on brand names, following several data-cleaning steps to ensure accuracy. Any missing or inconsistent codes were resolved in consultation with the study pharmacist (K.L.). “High-risk” medications were identified according to the 2019 Beers criteria (medications that are potentially inappropriate in most older adults, and medications that should be used with caution), which was the most updated version at that time [12], as well as from the literature [6,14,18,23]. Drugs were then grouped into 19 categories and coded as binary variables indicating whether they were taken. To ensure statistical power, only drug categories taken by at least 2% of participants were considered, resulting in 14 categories: NSAIDs, RAAS inhibitors, CCBs, diuretics, BZDs, sulfonylureas, proton pump inhibitors (PPIs), antithrombotics, antidepressants, peripheral alpha‐1 blockers, opioids, other central nervous system (CNS) drugs, antipsychotics and other anticholinergics (see S1 Table for further details).
Polypharmacy and hyper-polypharmacy were defined as the concomitant use of at least 5 or 10 drugs, respectively, including OTC drugs, vitamins and dietary supplements. This approach was chosen because, although commonly perceived as safe, these products can interact with prescription drugs, potentially altering their efficacy and safety [33,34]. Moreover, their inappropriate use (e.g., excessive dosing, prolonged use, or use for unsuitable indications) may also cause harm [35].
Outcome measures
The study file was linked to the national mortality registry held by the Israeli Ministry of Health using a unique identification number. Linkage was performed via TIMNA (Hebrew acronym for ‘Research Infrastructure for Big Data Studies’), a national research platform established by the Israeli government to enable big-data studies combining health data from multiple organizations. TIMNA personnel merged dates and causes of death, updated through 07 March 2022, into the study data file. They then removed all identifying fields, conducted thorough de-identification, including processing quasi-identifiers, and ensured compliance with k-anonymity standards. Following this process, we were granted accessed to the complete, de-identified dataset on 7 April 2022 at 11:30am for analysis. Cancer-related deaths were identified using relevant ICD-10 codes, which are used by the Israel Central Bureau of Statistics for official cause-of-death classification based on death notifications. For each participant, follow-up began at the date of examination and ended at death or the end of follow-up, whichever occurred first.
Potential confounders
The sociodemographic characteristics included age (continuous or age tertiles: < 69, 69–75, > 75 years), sex (men, women), years of education (continuous), ethnic origin (defined according to the reported country of birth, or the father’s country of birth for Israeli born individuals; categorized according to 4 main ethnic groups in Israel: Europe/America, Middle-East, North Africa, Yemen [36]), marital status (single, married, divorced/separated, widowed), occupation (categorized as Blue collar (manual and skilled trade occupations, including agriculture, fishing, craft, manufacturing, repair, construction, machine operation, transport, cleaning, packaging and general labor workers), lower White collar (administrative and clerical workers), and upper White collar (scientific and academic professionals, senior managers, freelancers and technical professionals)), and health maintenance organization (HMO) membership.
Lifestyle habits included leisure-time physical activity (any, none) and smoking status (current, past, never). Health factors included waist circumference (cm), body mass index (BMI, kg/m2; categorized according to the World Health Organization (WHO) criteria [37] as normal (<25), overweight (25–29.9) and obese (≥30)), self-rated health (excellent, very good, good, fair, poor), apolipoprotein E (ApoE) genotype (ApoE ε2 group includes ε2/2 and ε2/3; ApoE ε3 group includes ε3/3; ApoE ε4 group includes ε4/4 and ε3/4; and ‘Other’ group includes the rare ε2/4 or missing observations), and number of comorbidities (continuous). Information on chronic conditions was derived from the baseline interview. Participants were asked about diagnosed chronic conditions as well as past procedures and hospitalizations over the preceding 10 years. Conditions of interest were cancer, heart disease, lung disease, joint diseases, kidney diseases, diabetes, eye disease, gastrointestinal disease, and hypertension. This was supplemented by linkage with national hospitalization records, capturing all hospitalizations (diagnoses and procedures) prior to the baseline interview. Three conditions known as major mortality risk factors were considered individually: diabetes, cardiovascular disease and cancer.
Statistical analysis
Building upon the methodology proposed by Huang et al. [23], the analysis proceeded in two steps. First, we identified patterns of high-risk drug use in our cohort. Next, we examined whether any of these patterns were associated with an increased risk for all-cause or non-cancer mortality.
High-risk drug patterns.
Hierarchical agglomerative clustering with Ward’s method was carried out to group individuals into clusters containing similar profiles of the drug categories [23]. We selected hierarchical agglomerative clustering because it is well suited for small datasets, handles binary variables effectively [38], and is particularly appropriate when the objects being classified follow an underlying structured pattern [39], such as drug co-administration for specific conditions or comorbidities, as well as contraindications. To ensure cluster stability, only drug categories with a prevalence of 10% or more were considered in the clustering process. The optimal number of clusters was determined using dendrogram visualization, the Elbow plot (evaluating up to ten clusters), statistical power considerations for cluster sample sizes, and cluster stability as assessed by computing the mean Jaccard similarity index for each cluster over 1,000 repeated random subsets of the cohort using the clusterboot function from the fpc R package. A Jaccard index score > 0.5 indicates a reproducible and stable cluster [40,41]. The ultimate goal was to derive clear and clinically valuable drug patterns that exhibit differences in mortality [23,42]. Differences between clusters in baseline characteristics were tested using the Chi-Square test for categorical variables and the Kruskal–Wallis test for continuous variables.
Associations with mortality.
For both all-cause and non-cancer mortality, the number of events and unadjusted event rates per 1,000 person-years, along with 95% confidence intervals, were calculated overall and stratified by cluster.
Multivariable Cox proportional hazards regression models were used to examine associations of high-risk drug patterns with all-cause mortality. Since cancer incidence and mortality is rarely drug-related [43], an additional approach focused on death from non-cancer causes. Fine-Gray subdistribution hazard models were employed to account for cancer mortality as a competing event [44]. This method is based on the cumulative incidence function (CIF), which models the probability of the event of interest occurring in the presence of competing risks. That is, individuals who died from cancer (a competing event) remained in the risk set but were weighted differently to reflect that they could no longer experience the event of interest.
For each model, the proportional hazards assumption was assessed by conducting a test for the proportionality of the interaction between time and the variables in each model. If violations were identified, interactions between time and the relevant variables were incorporated into the model. All baseline variables were initially entered into the model, and removed sequentially using backward elimination, until only those with p < 0.10 or known confounders remained. Multicolinearity was evaluated by Cramér’s V statistic and IVF/Tolerance statistics. The Akaike’s information criterion (AIC) was used to select between alternative models.
Six sensitivity analyses were employed: (1) adjusting for individual cardiovascular-related diseases; (2) adjusting for pack years of smoking and time since quitting instead of smoking status; (3) considering only the pre-COVID-19 follow-up period, to address potential bias from pandemic-related changes in health-service utilization; (4) additionally adjusting for birth cohort to address time-related factors affecting drug exposure or its association with mortality; (5) considering an alternative definition for polypharmacy, excluding vitamins and dietary supplements; and (6) performing a multiple imputation analysis. Anthropometrics and ApoE genotyping were missing for up to 19.8% of participants, primarily because some completed the third follow-up by phone. We hypothesize that those who refused/could not arrive at the regional medical clinics were in poorer health, and therefore the data were missing not at random (MNAR). This is supported by the observation that participants with missing data were generally older, less physically active, less educated, had poorer self-rated health and a larger number of comorbidities (see S2 Table). In general, multiple imputation is not recommended when data are MNAR, as it can lead to biased estimates [45,46]. To minimize this risk and preserve sample size, we used the missing-indicator approach in the main analysis. Nevertheless, we conducted a sensitivity analysis using multiple imputation under the control-based pattern assumption [47]. Further details on the sensitivity analyses are provided in S1 Text.
Statistical analyses were conducted using SAS version 9.4 (SAS Institute, Cary, NC) and R version 4.2.2 (R Foundation for Statistical Computing) software.
Results
High-risk drug patterns
In our cohort of 1,048 community-dwelling older adults who were taking one or more medications, 814 (77.7%) took at least one high-risk drug. The most frequently used high-risk drugs were NSAIDs (reported by 43.0% of participants), followed by RAAS inhibitors (31.2%), CCBs (25.0%), diuretics (20.9%) and BZDs (17.3%). The prevalence of other high-risk drugs was less than 10% (S1 Table).
The resolution of five clusters (i.e., high-risk drug patterns) fit our data best (see S1 and S2 Figs for the dendrogram and Elbow plot, respectively): in Cluster 1 (n = 431) the most frequently taken drug, NSAIDs, was reported by less than one third (27.8%). This cluster also included all participants not taking any high-risk drugs, and was thus labeled ‘None’. Each of the other clusters had one drug category taken by all participants, which was used for labeling. Cluster 2 (‘CCBs’ cluster, n = 139) had a high prevalence of NSAIDs (55.4%, compared with 27.8%−54.6% in other clusters). Compared to the other clusters, there was also a high prevalence of peripheral alpha‐1 blockers (7.9%) and other CNS drugs (5.0%). Cluster 3 (‘RAAS inhibitors’ cluster, n = 174) had a high prevalence of sulfonylureas (13.2%, compared with 4.3%−9.1% in other clusters), and 54.6% of its participants were also taking NSAIDs. In Cluster 4 (‘Diuretics’ cluster, n = 140), 57.1% also took a CCB. This cluster also had a relatively high prevalence of antithrombotics (12.1%), PPIs (10.7%) and antipsychotics (5.7%). Cluster 5 (‘BZDs’ cluster, n = 164) also had a high prevalence of antidepressant use (10.4%, compared to 5.1%−7.9% in other clusters). The prevalence of all 14 high-risk drug categories across clusters is presented in Fig 1.
Alpha-1 blockers, Peripheral alpha-1 blockers unless for benign prostatic hyperplasia; BZDs, Benzodiazepines (including ‘Z-drugs’); CCBs, Calcium channel blockers; NSAIDs, Nonsteroidal anti-inflammatory drugs; PPIs, Proton-pump inhibitors (use for >8 weeks unless high‐risk patients); RAAS, Renin angiotensin-aldosterone system. Cluster 1 = ‘None’ cluster (n = 431); Cluster 2 = ‘Calcium channel blockers (CCBs)’ cluster (n = 139); Cluster 3 = ‘Renin-angiotensin-aldosterone system (RAAS) inhibitors’ cluster (n = 174); Cluster 4 = ‘Diuretics’ cluster (n = 140); Cluster 5 = ‘Benzodiazepines (BZDs)’ cluster (n = 164).
Table 1 presents baseline characteristics of the participants overall and by cluster. The mean age at baseline was 73.3 ± 7.3 years, with 56% women. The ‘None’ cluster included the youngest participants (mean age 71.6 ± 7.2), had the lowest mean BMI (27.2 ± 4.2), and the highest proportion engaged in physical activity (54.8%). Its participants also had the fewest comorbidities (a mean of 2.0 ± 1.5), the highest self-rated excellent health (11.1%, compared with 1.5%−5.4% in other clusters), and only 17.4% were exposed to polypharmacy. The ‘CCBs’ cluster had the highest prevalence of hypertension (90.6%) and cardiovascular disease (61.2%). The ‘RAAS inhibitors’ cluster was the only cluster with a majority of men (54.0%), and had the highest prevalence of diabetes (39.7%). The ‘Diuretics’ cluster had the oldest participants (mean age 76.3 ± 7.1), the highest proportion of participants reporting fair/poor self-rated health (39.0%), not engaging in any leisure-time physical activity (61.4%), and the highest mean number of comorbidities (3.8 ± 1.8), more specifically joint disease (34.3%), lung disease (25.0) and kidney disease (22.9%). This cluster also had the highest prevalence of polypharmacy (75.7%). The ‘BZDs’ cluster included relatively older participants (mean age 75.0 ± 7.0), with nearly two-thirds women. This cluster had the highest prevalence of gastrointestinal disease (53.0%) across clusters, and the second-highest polypharmacy prevalence (67.7%). Sociodemographic characteristics were similar across the clusters.
High-risk drug patterns and mortality
Overall, 64.8% of participants died over a median follow-up of 14.3 years (13,591 person-years). The number of events and unadjusted event rates for all-cause and non-cancer mortality, both overall and by cluster, are presented in Table 2.
The ‘None’ cluster, which was the largest and had no prevalent high-risk drug, was set as the reference category. In the unadjusted Cox regression model (see Table 3), compared with the ‘None’ cluster, each of the other drug patterns was associated with increased all-cause mortality risk (HRs range 1.35–2.19). Interestingly, adjustment for age and sex alone substantially attenuated the observed associations (S3 Table). In the fully adjusted model, considering sociodemographic, lifestyle, genetic, subjective and objective health measures, only the ‘Diuretics’ cluster was significantly associated with increased mortality risk (HR = 1.33, 95%CI: 1.03–1.72). The ‘CCBs’ cluster had an HR of 1.27 (95%CI: 0.98–1.64), whereas the other clusters presented no increased risk (HRs = 0.98 and 1.05 for the ‘RAAS inhibitors’ and ‘BZDs’ clusters, respectively; Table 3 and S4 Table). The adjusted survival curves for the five high-risk drug patterns are presented in Fig 2.
Cox proportional hazard model was adjusted for: age, sex, polypharmacy, physical activity, BMI, smoking, ApoE genotype, self-rated health, diabetes, cardiovascular disease, cancer and count of remaining comorbidities. High-risk drug patterns: Cluster 1 = ‘None’ cluster (reference); Cluster 2 = ‘Calcium channel blockers (CCBs)’ cluster; Cluster 3 = ‘Renin-angiotensin-aldosterone system (RAAS) inhibitors’ cluster; Cluster 4 = ‘Diuretics’ cluster. Only Cluster 4 significantly differed from Cluster 1 (HR = 1.33, 95%CI: 1.03-1.72); Cluster 5 = ‘Benzodiazepines (BZDs)’ cluster.
In the competing risks analysis, both the ‘Diuretics’ and ‘CCBs’ clusters were significantly associated with increased non-cancer mortality risk (HR = 1.41 and 1.47, respectively), after adjusting for sociodemographic, lifestyle, genetic, and both subjective and objective health measures, and accounting for violations of the proportional hazards assumption (see Table 3 and S5 Table). The ‘RAAS inhibitors’ cluster was also associated with some increased risk, although to a lesser extent (HR = 1.25, 95%CI: 0.97–1.61), whereas the ‘BZDs’ cluster did not differ from the ‘None’ cluster.
Results remained similar across the sensitivity analyses (see Models 1–6 in Table 4).
Discussion
Our study revealed five distinct high-risk drug patterns among community-dwelling older adults: (1) ‘None’ cluster (i.e., no dominant high-risk drug), (2) ‘CCBs’ cluster, (3) ‘RAAS inhibitors’ cluster, (4) ‘Diuretics’ cluster and (5) ‘BZDs’ cluster. The ‘Diuretics’ cluster represented the most complex patients (i.e., were older, had a higher number of comorbidities and a higher prevalence of polypharmacy), and was most consistently associated with increased mortality risk, for both all-cause and non-cancer mortality. This suggests that older adults taking diuretics, PPIs, antithrombotics and antipsychotics face increased mortality risk and warrant special attention. An increased risk, mainly for non-cancer mortality, was also observed for the ‘CCBs’ cluster, whereas the other clusters did not significantly differ from the ‘None’ cluster in any of the analyses. Findings from the competing risks analysis were directionally consistent with the Cox models, with slightly stronger effect estimates, supporting our assumption that cancer deaths are less likely to be influenced by drug use.
Several drug combinations could have contributed to the increased risk observed for the ‘Diuretics’ cluster. Over 57% of its participants were concurrently using a CCB. Although their concomitant use may be appropriate in cases requiring multiple agents or with an underlying cardiovascular disease, it nevertheless might exacerbate the risk of ADRs [48]. Moreover, in some cases this may result from an inappropriate CCB-diuretic prescribing cascade, which raises the risk of falls, acute kidney injury, and other serious adverse events [18,49]. Older hypertensive patients should undergo regular medication review to ensure optimal treatment. Clinicians should be aware of nonpharmacologic strategies for CCB-induced edema.
Over half of participants in the ‘Diuretics’ cluster also used NSAIDs, which can reduce diuretic efficacy. Additionally, drug interactions can exacerbate NSAIDs-related ADRs, such as bleeding, coronary events and renal insufficiency [50], a concern given the high prevalence of kidney disease in this cluster. Toxicity related to the concomitant use of NSAIDs and antihypertensives may also have contributed to the increased non-cancer mortality risk seen in the ‘CCBs’ cluster [12,15].
The ‘Diuretics’ cluster had the highest prevalence of PPIs and antithrombotics. In addition to their individual risks (e.g., Clostridium difficile infection and fractures from prolonged PPI use [12], and increased thrombosis and bleeding risks with antithrombotics [51]), their combined use can cause harmful interactions and reduce antithrombotic efficacy [52]. Finally, this cluster had the highest antipsychotic use. While crucial for some conditions, their wide range of side effects [12] might reduce quality of life and impair medication adherence [53]. Patient preferences should guide antipsychotic choices.
Of note are the differences in sex distribution across clusters. Only the ‘RAAS inhibitors’ cluster had a majority of men. Moreover, Cluster 5, with the highest BZD and antidepressant use, had nearly two-thirds women. A similar pattern was observed by Huang et al. [23], where the cluster with the highest prevalence of RAAS inhibitors was the only one with a majority of men, and the ‘Mental health drugs’ cluster had 61.1% women. Future studies should explore sex-specific high-risk drug patterns in older people.
Interestingly, even after adjustment for socioeconomic, lifestyle, and health-related measures, ApoE genotype was significantly associated with a 1.5-fold increased mortality risk. ApoE genotype may affect both exposure (via predisposition to certain diseases [54,55]) and mortality risk [56]. Furthermore, pharmacogenomic/pharmacogenetic studies suggest a role for ApoE polymorphism in the inter-individual variability in drug responses, with ε4 carriers having poor responses to certain treatments [57,58] and higher ADR risks [58,59]. The mechanism may involve effects of the ApoE gene on liver drug metabolism [57]. Owing to limited statistical power, we could not explore possible interactions between ApoE genotype and drug patterns. Further research is needed to investigate ApoE as a biomarker for pharmacogenetic responses.
In addition, the lack of variation in education and occupation across medication clusters may reflect the influence of Israel’s universal health insurance system, which ensures access to essential medications and healthcare services for all citizens. However, not all medications are covered under the subsidized Health Basket, and many require co-payments. While certain protections exist for vulnerable populations, such as quarterly spending caps for chronically ill patients, welfare recipients and holocaust survivors, research has shown that pharmaceutical co-payments can still pose a substantial financial burden on the poor, elderly, and chronically ill, potentially impairing adherence [60]. Sociodemographic factors should be considered when analyzing medication use and health outcomes, even in a universal healthcare setting.
Finally, polypharmacy was not significantly associated with mortality risk in any of the analyses. This suggests that considering medication patterns may be more valuable in identifying high-risk patients than merely their count.
Several study limitations should be acknowledged. First, medication data were collected only at baseline and may not reflect current prescribing trends or changes over time affecting drug exposure or its association with mortality (such as shifts in population health behaviors). Participants may also have altered their drug use during follow-up due to health changes or adherence issues. The broad medication categories, encompassing one or more drug classes, partly addressed this limitation. We also conducted two sensitivity analyses to account for time-related factors (a potential cohort effect and possible pandemic-related changes in health-services utilization or health behaviors), both of which yielded consistent results, supporting the robustness of our findings. Second, although we adjusted for multiple health-related factors, we did not account for dosage or adherence, which could influence drug effects and interactions. Third, the sample size limited our ability to assess additional cause-specific mortality outcomes. Fourth, some self-reported measures (e.g., physical activity, smoking history) may be subject to recall bias. Fifth, our cohort included only Jewish older adults in Israel, which may limit generalizability, although participants were from diverse ethnic backgrounds. Sixth, survival bias due to informative censoring and/or left truncation is possible, as participants who died before their third interview were not included. This may have led to underestimating high-risk drug use in the general older population, but offered insights into the impact on relatively healthy older adults. Finally, our study was not designed to establish causal inference about specific drugs, but rather to identify profiles of high-risk patients, defined by their patterns of drug use, who may benefit from interventions targeting their drug regimens. We therefore considered only potentially inappropriate or high-risk medications, investigated data-driven patterns rather than pre-defined combinations or individual drugs, and examined their associations with mortality while adjusting for a broad range of potential confounders, including sociodemographic and health-related characteristics. These included individual conditions known as major risk factors of mortality, a count of additional chronic conditions, and self-rated health, a measure consistently associated with mortality and reflective of severity of both diagnosed and undiagnosed diseases [61].
The strengths of this study include its population-based, ethnically diverse sample of community-dwelling older men and women, the high-quality data, recorded at a state of health for research purposes, along with the availability of information on OTC drugs, which enabled us to identify ‘real-world’ high-risk drug patterns. Furthermore, the complete long-term follow-up enabled the accumulation of ADR-related downstream clinical outcomes, ultimately leading to premature mortality. Finally, our study employs a patient-centered, advanced approach, combining hierarchical clustering with survival analysis, to identify clinically meaningful usage patterns of common high-risk medications among community-dwelling older adults.
Conclusion
In this prospective cohort study of community-dwelling older adults, five clusters were identified. The concurrent use of diuretics, PPIs, antithrombotics and antipsychotics (‘Diuretics’ cluster) was most consistently associated with overall and non-cancer mortality. The ‘CCBs’ cluster showed a greater risk mainly for non-cancer mortality. Specific high-risk drug combinations, such as NSAIDs with antihypertensives, may have further contributed to the observed increased mortality risk. As patterns of drug regimens reflect disease burden and treatment complexity, may confer additional risks (including through drug-disease interactions), and are modifiable, they may serve as a practical tool to support risk stratification efforts and guide early interventions for vulnerable older patients (e.g., frequent medication reviews, dosage adjustments and closer monitoring of specific biomarkers/ADRs). To advance more personalized care, future research can harness artificial intelligence to characterize high-risk drug patterns across diverse subpopulations.
Supporting information
S1 Text. Detailed information on methodology.
https://doi.org/10.1371/journal.pone.0332210.s001
(DOCX)
S1 Table. High-risk drug categories used by 1,048 community-dwelling older adults.
https://doi.org/10.1371/journal.pone.0332210.s002
(DOCX)
S2 Table. Selected baseline characteristics of study participants by completeness of data.
https://doi.org/10.1371/journal.pone.0332210.s003
(DOCX)
S3 Table. Cox proportional hazard models for the association between high-risk drug patterns and all-cause mortality, with incremental adjustment.
https://doi.org/10.1371/journal.pone.0332210.s004
(DOCX)
S4 Table. Fully adjusted Cox proportional hazard model for the association between high-risk drug patterns and all-cause mortality.
https://doi.org/10.1371/journal.pone.0332210.s005
(DOCX)
S5 Table. Adjusted Fine and Gray subdistribution hazard model for the association between high-risk drug patterns and non-cancer mortality.
https://doi.org/10.1371/journal.pone.0332210.s006
(DOCX)
S1 Fig. Dendrogram of cluster analysis in 1,048 community-dwelling older adults.
https://doi.org/10.1371/journal.pone.0332210.s007
(PDF)
S2 Fig. Elbow method plot for determining the optimal number of clusters.
https://doi.org/10.1371/journal.pone.0332210.s008
(PDF)
Acknowledgments
This work was performed in partial fulfillment of the requirements for a PhD degree of Liat Orenstein, Gray Faculty of Medical and Health Sciences, Tel Aviv University, Tel Aviv, Israel.
References
- 1. Souza DLB, Oliveras-Fabregas A, Minobes-Molina E, de Camargo Cancela M, Galbany-Estragués P, Jerez-Roig J. Trends of multimorbidity in 15 European countries: a population-based study in community-dwelling adults aged 50 and over. BMC Public Health. 2021;21(1):76. pmid:33413239
- 2. Boyd CM, Darer J, Boult C, Fried LP, Boult L, Wu AW. Clinical practice guidelines and quality of care for older patients with multiple comorbid diseases: implications for pay for performance. JAMA. 2005;294(6):716–24. pmid:16091574
- 3. Bourgeois FT, Orenstein L, Ballakur S, Mandl KD, Ioannidis JPA. Exclusion of elderly people from randomized clinical trials of drugs for ischemic heart disease. J Am Geriatr Soc. 2017;65(11):2354–61. pmid:28306144
- 4. Bayer A, Tadd W. Unjustified exclusion of elderly people from studies submitted to research ethics committee for approval: descriptive study. BMJ. 2000;321(7267):992–3. pmid:11039965
- 5. Van Spall HGC, Toren A, Kiss A, Fowler RA. Eligibility criteria of randomized controlled trials published in high-impact general medical journals: a systematic sampling review. JAMA. 2007;297(11):1233–40. pmid:17374817
- 6. Davies EA, O’Mahony MS. Adverse drug reactions in special populations - the elderly. Br J Clin Pharmacol. 2015;80(4):796–807. pmid:25619317
- 7. Anathhanam S, Powis RA, Cracknell AL, Robson J. Impact of prescribed medications on patient safety in older people. Ther Adv Drug Saf. 2012;3(4):165–74. pmid:25083234
- 8. Corsonello A, Pedone C, Incalzi RA. Age-related pharmacokinetic and pharmacodynamic changes and related risk of adverse drug reactions. Curr Med Chem. 2010;17(6):571–84. pmid:20015034
- 9. Beers MH, Ouslander JG, Fingold SF, Morgenstern H, Reuben DB, Rogers W, et al. Inappropriate medication prescribing in skilled-nursing facilities. Ann Intern Med. 1992;117(8):684–9. pmid:1308759
- 10. American Geriatrics Society 2012 Beers Criteria Update Expert Panel. American Geriatrics Society updated Beers Criteria for potentially inappropriate medication use in older adults. J Am Geriatr Soc. 2012;60(4):616–31. pmid:22376048
- 11. By the American Geriatrics Society 2015 Beers Criteria Update Expert Panel. American Geriatrics Society 2015 updated beers criteria for potentially inappropriate medication use in older adults. J Am Geriatr Soc. 2015;63(11):2227–46. pmid:26446832
- 12. By the 2019 American Geriatrics Society Beers Criteria® Update Expert Panel. American Geriatrics society 2019 updated AGS Beers Criteria® for potentially inappropriate medication use in older adults. J Am Geriatr Soc. 2019;67(4):674–94. pmid:30693946
- 13. By the 2023 American Geriatrics Society Beers Criteria® Update Expert Panel. American Geriatrics Society 2023 updated AGS Beers Criteria® for potentially inappropriate medication use in older adults. J Am Geriatr Soc. 2023;71(7):2052–81. pmid:37139824
- 14. Lin C, Wen Y, Chen L, Hsiao F. Potentially high-risk medication categories and unplanned hospitalizations: a case-time-control study. Sci Rep. 2017;7(1):41035.
- 15. Makris UE, Pugh MJ, Alvarez CA, Berlowitz DR, Turner BJ, Aung K, et al. Exposure to high-risk medications is associated with worse outcomes in older veterans with chronic pain. Am J Med Sci. 2015;350(4):279–85. pmid:26418380
- 16. Wojt IR, Cairns R, Clough AJ, Tan ECK. The prevalence and characteristics of psychotropic-related hospitalizations in older people: a systematic review and meta-analysis. J Am Med Dir Assoc. 2021;22(6):1206–14.e5. pmid:33539820
- 17. Oscanoa TJ, Lizaraso F, Carvajal A. Hospital admissions due to adverse drug reactions in the elderly. A meta-analysis. Eur J Clin Pharmacol. 2017;73(6):759–70. pmid:28251277
- 18. Rochon PA, Austin PC, Normand S-L, Savage RD, Read SH, McCarthy LM, et al. Association of a calcium channel blocker and diuretic prescribing cascade with adverse events: a population-based cohort study. J Am Geriatr Soc. 2024;72(2):467–78. pmid:38009803
- 19. Lombardi N, Bettiol A, Crescioli G, Ravaldi C, Bonaiuti R, Venegoni M, et al. Risk of hospitalisation associated with benzodiazepines and z-drugs in Italy: a nationwide multicentre study in emergency departments. Intern Emerg Med. 2020;15(7):1291–302. pmid:32333265
- 20. Price SD, Holman CDJ, Sanfilippo FM, Emery JD. Impact of specific Beers Criteria medications on associations between drug exposure and unplanned hospitalisation in elderly patients taking high-risk drugs: a case-time-control study in Western Australia. Drugs Aging. 2014;31(4):311–25. pmid:24615204
- 21. Hanlon P, Quinn TJ, Gallacher KI, Myint PK, Jani BD, Nicholl BI, et al. Assessing risks of polypharmacy involving medications with anticholinergic properties. Ann Fam Med. 2020;18(2):148–55. pmid:32152019
- 22. Ruxton K, Woodman RJ, Mangoni AA. Drugs with anticholinergic effects and cognitive impairment, falls and all-cause mortality in older adults: a systematic review and meta-analysis. Br J Clin Pharmacol. 2015;80(2):209–20. pmid:25735839
- 23. Huang Y-T, Steptoe A, Wei L, Zaninotto P. The impact of high-risk medications on mortality risk among older adults with polypharmacy: evidence from the English Longitudinal Study of Ageing. BMC Med. 2021;19(1):321. pmid:34911547
- 24. Masnoon N, Kalisch Ellett L, Shakib S, Caughey GE. Predictors of mortality in the older population: the role of polypharmacy and other medication and chronic disease-related factors. Drugs Aging. 2020;37(10):767–76. pmid:32885396
- 25. Alexander M, Kiang M, Barbieri M. Trends in black and white opioid mortality in the United States, 1979–2015. Epidemiology. 2018;29(5):707–15.
- 26. Patorno E, Glynn RJ, Levin R, Lee MP, Huybrechts KF. Benzodiazepines and risk of all cause mortality in adults: cohort study. BMJ. 2017;358:j2941. pmid:28684397
- 27. Glynn RJ, Knight EL, Levin R, Avorn J. Paradoxical relations of drug treatment with mortality in older persons. Epidemiology. 2001;12(6):682–9. pmid:11679797
- 28. Menichelli D, Poli D, Antonucci E, Palareti G, Pignatelli P, Pastori D, et al. Renin-angiotensin-aldosterone system inhibitors and mortality risk in elderly patients with atrial fibrillation. Insights from the nationwide START registry. Eur J Intern Med. 2024;119:84–92. pmid:37648584
- 29. Fahmi A, Wong D, Walker L, Buchan I, Pirmohamed M, Sharma A, et al. Combinations of medicines in patients with polypharmacy aged 65-100 in primary care: large variability in risks of adverse drug related and emergency hospital admissions. PLoS One. 2023;18(2):e0281466. pmid:36753492
- 30. Anoopkumar-Dukie S, Mey A, Hall S, Bernaitis N, Davey AK, Plummer D. Non-prescription medicines may contribute to non-adherence to prescription medicines in people living with chronic health conditions. Int J Clin Pract. 2020;74(6):e13489. pmid:32083362
- 31. Schmiedl S, Rottenkolber M, Hasford J, Rottenkolber D, Farker K, Drewelow B, et al. Self-medication with over-the-counter and prescribed drugs causing adverse-drug-reaction-related hospital admissions: results of a prospective, long-term multi-centre study. Drug Saf. 2014;37(4):225–35. pmid:24550104
- 32. Abdu N, Mosazghi A, Teweldemedhin S, Asfaha L, Teshale M, Kibreab M, et al. Non-Steroidal Anti-Inflammatory Drugs (NSAIDs): usage and co-prescription with other potentially interacting drugs in elderly: a cross-sectional study. PLoS One. 2020;15(10):e0238868. pmid:33035226
- 33. Mielke N, Huscher D, Douros A, Ebert N, Gaedeke J, van der Giet M, et al. Self-reported medication in community-dwelling older adults in Germany: results from the Berlin Initiative Study. BMC Geriatr. 2020;20(1):22. pmid:31964342
- 34. Disdier Moulder MPA, Hendricks AK, Ou NN. Towards appropriate polypharmacy in older cardiovascular patients: How many medications do I have to take? Clin Cardiol. 2020;43(2):137–44. pmid:31825133
- 35. Phua DH, Zosel A, Heard K. Dietary supplements and herbal medicine toxicities-when to anticipate them and how to manage them. Int J Emerg Med. 2009;2(2):69–76. pmid:20157447
- 36. Dankner R, Shanik M, Roth J, Luski A, Lubin F, Chetrit A. Sex and ethnic-origin specific BMI cut points improve prediction of 40-year mortality: the Israel GOH study. Diabetes Metab Res Rev. 2015;31(5):530–6. pmid:25689480
- 37.
World Health Organization. Obesity: preventing and managing the global epidemic. Geneva: World Health Organization; 2000.
- 38. Gupta A, Sharma H, Akhtar A. A comparative analysis of k-means and hierarchical clustering. EPRA Int J Multidiscip Res (IJMR). 2021;7(8):412–8.
- 39. Cornell J, Pugh JA, Williams JWJ, Kazis L, Lee AF, Parchman ML. Multimorbidity clusters: clustering binary data from a large administrative medical database. Appl Multivar Res. 2008;12(3):163–82.
- 40. Hennig C. Dissolution point and isolation robustness: robustness criteria for general cluster analysis methods. J Multivar Anal. 2008;99(6):1154–76.
- 41. Robert M, Jouanjus E, Khouri C, Fouilhé Sam-Laï N, Revol B. The opioid epidemic: a worldwide exploratory study using the WHO pharmacovigilance database. Addiction. 2023;118(4):771–5. pmid:36331523
- 42. Teh RO, Menzies OH, Connolly MJ, Doughty RN, Wilkinson TJ, Pillai A, et al. Patterns of multi-morbidity and prediction of hospitalisation and all-cause mortality in advanced age. Age Ageing. 2018;47(2):261–8. pmid:29281041
- 43. Schöttker B, Saum K, Muhlack D, Hoppe L, Holleczek B, Brenner H. Polypharmacy and mortality: new insights from a large cohort of older adults by detection of effect modification by multi-morbidity and comprehensive correction of confounding by indication. Eur J Clin Pharmacol. 2017;73:1041–8.
- 44. Fine J, Gray R. A proportional hazards model for the subdistribution of a competing risk. J Am Stat Assoc. 1999;94(446):496.
- 45. Yang S, Wang L, Ding P. Causal inference with confounders missing not at random. Biometrika. 2019;106(4):875–88.
- 46. Rubin DB. Inference and missing data. Biometrika. 1976;63(3):581–92.
- 47. Li L. SAS® V9.4 MNAR statement for multiple imputations for missing not at random in longitudinal clinical trials. PharmaSUG 2019 Conference. 2019.
- 48. Peeters L, Kester MP, Feyz L, Van Den Bemt PMLA, Koch BCP, Van Gelder T. Pharmacokinetic and pharmacodynamic considerations in the treatment of the elderly patient with hypertension. Expert Opin Drug Metab Toxicol. 2019;4:287–97.
- 49. Savage RD, Visentin JD, Bronskill SE, Wang X, Gruneir A, Giannakeas V, et al. Evaluation of a common prescribing cascade of calcium channel blockers and diuretics in older adults with hypertension. JAMA Intern Med. 2020;180(5):643–51. pmid:32091538
- 50. Moore N, Pollack C, Butkerait P. Adverse drug reactions and drug-drug interactions with over-the-counter NSAIDs. Ther Clin Risk Manag. 2015;11:1061–75. pmid:26203254
- 51. Arahata M, Asakura H. Antithrombotic therapies for elderly patients: handling problems originating from their comorbidities. Clin Interv Aging. 2018;13:1675–90. pmid:30237704
- 52. Cena C, Traina S, Parola B, Bo M, Fagiano R, Siviero C. Prescription of proton pump inhibitors in older adults with complex polytherapy. Eur J Hosp Pharm. 2020;27(6):341–5. pmid:33097617
- 53. Pillinger T, Howes OD, Correll CU, Leucht S, Huhn M, Schneider-Thoma J, et al. Antidepressant and antipsychotic side-effects and personalised prescribing: a systematic review and digital tool development. Lancet Psychiatry. 2023;10(11):860–76. pmid:37774723
- 54. Song Y, Stampfer MJ, Liu S. Meta-analysis: apolipoprotein E genotypes and risk for coronary heart disease. Ann Intern Med. 2004;141(2):137–47. pmid:15262670
- 55. Raber J, Huang Y, Ashford JW. ApoE genotype accounts for the vast majority of AD risk and AD pathology. Neurobiol Aging. 2004;25(5):641–50. pmid:15172743
- 56. Deelen J, Beekman M, Uh HW, Broer L, Ayers KL, Tan Q, et al. Genome-wide association meta-analysis of human longevity identifies a novel locus conferring survival beyond 90 years of age. Hum Mol Genet. 2014;23(16):4420–32.
- 57. Cacabelos R. Pharmacogenomics and therapeutic prospects in dementia. Eur Arch Psychiatry Clin Neurosci. 2008;258 Suppl 1:28–47. pmid:18344047
- 58. Wei C-Y, Lee M-TM, Chen Y-T. Pharmacogenomics of adverse drug reactions: implementing personalized medicine. Hum Mol Genet. 2012;21(R1):R58-65. pmid:22907657
- 59. Nebes RD, Pollock BG, Perera S, Halligan EM, Saxton JA. The greater sensitivity of elderly APOE ε4 carriers to anticholinergic medications is independent of cerebrovascular disease risk. Am J Geriatr Pharmacother. 2012;10(3):185–92. pmid:22534472
- 60. Simon-Tuval T, Triki N, Chodick G, Greenberg D. Determinants of cost-related nonadherence to medications among chronically Ill patients in Maccabi Healthcare Services, Israel. Value Health Reg Issues. 2014;4:41–6. pmid:29702805
- 61. Lima-Costa MF, Cesar CC, Chor D, Proietti FA. Self-rated health compared with objectively measured health status as a tool for mortality risk screening in older adults: 10-year follow-up of the Bambuí Cohort Study of Aging. Am J Epidemiol. 2012;175(3):228–35. pmid:22193172