Number of medically prescribed pharmaceutical agents as predictor of mortality risk: a longitudinal, time-variable analysis in the EPIC-Heidelberg cohort

The number of prescribed medications might be used as proxy indicator for general health status, in models to predict mortality risk. To estimate the time-varying association between active pharmaceutical ingredient (API) count and all-cause mortality, we analyzed data from a population cohort in Heidelberg (Germany), including 25,546 participants with information on medication use collected at 3-yearly intervals from baseline recruitment (1994–1998) until end of 2014. A total of 4548 deaths were recorded until May 2019. Time-dependent modeling was used to estimate hazard ratios (HR) and their 95% confidence intervals (CI) for all-cause mortality in relation to number of APIs used, within three strata of age (≤ 60, > 60 to  ≤ 70 and > 70 years) and adjusting for lifestyle-related risk factors. For participants reporting commonly used APIs only (i.e., API types accounting for up to 80% of medication time in the population) total API counts showed no association with mortality risk within any age stratum. However, when at least one of the APIs was less common, the total API count showed a strong relationship with all-cause mortality especially up to age ≤ 60, with HR up to 3.70 (95% CI 2.30–5.94) with 5 or 6 medications and 8.19 (5.61–11.97) for 7 or more APIs (versus none). Between > 60 and 70 years of age this risk association was weaker, with HR up to 3.96 (3.14–4.98) for 7 or more APIs, and above 70 years it was weakened further (HR up to 1.54 (1.34–1.79)). Multiple API-use may predict mortality risk in middle-aged and women and men ≤ 70 years, but only if it includes at least one less frequently used API type. With advancing age, and multiple medication becomes increasingly prevalent, the association of API count with risk of death progressively attenuates, suggesting an increasing complexity with age of underlying mortality determinants.

The statistical analyses were restricted to APIs classified at level 2 of ATC coding (e.g., "A10", for all glucoselowering medications).The total number of APIs used by each individual was updated at each of the six follow-up rounds after baseline.We categorized an individual as being in the polypharmacy group if he or she was taking 5 or more API in order to compare our results to previous investigations.To determine which API individuals had been taking most frequently and for the longest duration we tabulated the total person-years spent on API across age strata (age windows of ≤ 60, 60-≤ 70, and > 70 years).Out of 25,520 individuals, 19,677 (77% of the cohort) had complete medication information at baseline and across all the six follow-up questionnaires (FUP1, FUP2, …., FUP6), 12% missed one, 6% missed two and 5% had missing data on medication use on three or more follow-up occasions.The numbers of missing values were equally distributed across the three age windows.In our statistical modeling, from baseline till FUP6, the last observation on medication use was carried forward up until either the subsequent follow-up ascertainment of medical use or, if a participant did not respond to a follow-up questionnaire, for 3 further years (or date of death, whatever came first).For all participants who had completed the 6th follow-up questionnaire (FUP6), this last observation was carried forward until the final censoring date of May 2019 (or date of death, whichever came first).The median duration between FUP6 and date of censoring, or death, was 5.5 years (inter-quartile range, 5.3-5.7 years).
All proportional hazards analyses were adjusted for age, which is implicit in our left-truncated age-as-timescale model.In addition, models were adjusted for several baseline-covariates, including body mass index (BMI), physical activity, smoking intensity and history [lifetime duration of smoking (years), average daily cigarettes, and time since quitting (in years) for former smokers], and highest level of formal education as a proxy for socioeconomic status.We used Schoenfeld residuals and the "cox.zph"function of the "survival" package in R-studio to determine possible violations of the proportional hazards assumption for each variable.The proportional hazards assumption appeared violated for mortality risk in relation to the number of APIs, indicating that the relative risk association of API count with mortality was not constant with age; we therefore allowed the hazard ratio for API to change with age, and estimated separate hazard ratios for each of three age windows: below age 60, between 60 and 70 years, and 70 years and above.We also tested for interactions between API count and sex regarding mortality risks, but we found no evidence for such interaction.
Since the concept of multiple medication use such as the polypharmacy score or other, combines multiple medications of differing clinical effect, we investigated which medications were responsible for the association of multiple medication-use with mortality, by building a similar score using only the APIs responsible for 80% of the person-years of API use by the participants (further referred to as "common" APIs), and comparing the effect of this score to one constructed using the remaining 20% of API use, further referred to as "uncommon" APIs, Appendix 2.
All statistical analyses were carried out using R (version 4.0.1,2020) 21 .

Ethics declaration
This project is covered by the ethical approval for the EPIC-Heidelberg cohort (Ethical Committee of the Medical Faculty Heidelberg, reference number 13/94).

Consent for participation
Informed consent was obtained from all participants at baseline (1994-1998).

Results
Baseline characteristics of the study population (N = 25,540) are displayed in Table 1.A total of 4,548 deaths (18% or study participants) were recorded, over an average of about 20 years of prospective follow-up.The mean ages at recruitment were 50 years (SD = 8 years, interquartile range (IQR) = 43-56 years) and 56 years (SD = 7 years, IQR = 53-62 years), and the mean follow-up durations were 22 years (SD = 2 years, IQR = 21-23 years) and 14 years (SD = 6 years, 10-19 years), respectively, for the participants alive at the end of the study and for those deceased.Although men represented only 43% of the study participants, the proportion of men in the deceased group was 64%.Furthermore, thepercentage of current smokers (32%) was higher in the deceased group than in the alive group (21%).The prevalence of polypharmacy was more than three times higher in the deceased group (12%) compared to the alive group (4%), and the prevalence for self-reported diseases at baseline recruitment was also higher in the deceased group.After excluding four ATC code groups (Appendix 1), 80 percent of medication use time was related to only a relatively small subset of medication classes (Appendix 2A; two-level ATC categories), which included only 16 basic ATC medication classes up to the age of 60, and only 12 basic ATC medication groups when EPIC study participants were older than 70.The most common prescriptions were for medications targeting the cardiovascular system (C-category), which represented 33.4% of medication use up to age 60, increasing to 47.4% when participants were older than 70.The use of anti-thrombotic medications (B01) gradually rose from 4.8% of time below age 60 to 11.9% after age 70, whereas the relative part of medications used for thyroid diseases diminished from 11.1% to 4.6%.Other categories of commonly used APIs included glucose-lowering medications (A10, 2.4-3.9%depending on age window), medications used to treat gastro-intestinal reflux problems (A02, 3.7-4.3%)and APIs with anti-inflammatory and/or anti-rheumatic actions (M01, 4.7-2.7%).The group of less commonly used medications, which jointly accounted for the remaining 20% of API use, included a total of 29 API classes described in Appendix 2B, where each individual class accounted for only 0.3% to 2.5% of estimated time of use (Appendix 2A).The numbers of APIs used simultaneously, increased with age, almost linearly (Fig. 1).
Models adjusted for age, BMI, physical activity, smoking history, highest school level as a proxy for socioeconomic status, and self-reported myocardial infraction, diabetes, stroke, and hypertension at baseline recruitment showed significant associations between increasing API count and mortality rates (Table 2).Yet, the strength of www.nature.com/scientificreports/association between API count and risk of death varied across different age windows.Below the age of 60, we observed a hazard ratio (HR) of 1.87 (95% CI: 1.20, 2.91) for those using 4 APIs, 2.50 (1.67-3.73)for those using 5 or 6 APIs and 6.28 (95% CI: 4.37, 9.03) for those using 7 or more APIs, relative to individuals who did not use any medications.For the age window of 60-70 years the HR estimates were moderately lower.For 70 years and above, the number of APIs used overall had no major association with mortality risk, except for individuals using seven or more APIs [HR = 1.35 (95% CI: 1.17, 1.55)].
Comparing commonly used to less used medications, we observed that the relative mortality hazards were increased especially for those individuals who used at least one uncommon API type (Table 2) (Fig. 2).For individuals aged 60 years or younger, the HR for five or six APIs versus none is 1.39 (95% CI: 0.76, 2.56), when no uncommon APIS are present, and 1.65 (95% CI: 0.67-4.07)for those using 7 or more APIs.However, when at least one of the APIs was of a less commonly used type the HRs increased to 3.70 (95% CI: 2.30, 5.94) and 8.19 (95% CI: 5.61-11.97),respectively.Also, in the age window of > 60 to ≤ 70 years an increasing API count was related to increasing mortality risk only when medication use included at least one less common API type, but with more moderate HRs as compared to the younger age window.Above age 70, even when at least one less common API type was used, higher API counts were not related to any major increase in mortality risk unless 7 or more APIs were used (HR = 1.54; 95%CI: 1.34-1.79).

Discussion
Using a time-dependent modelling approach we assessed the relationships between age-varying medication use and all-cause mortality in a general population cohort of men and women initially 35 to 65 years of age (40-65 years for men).Adjusting for age, smoking history, BMI and level of formal education we found that mortality risk increased significantly with increasing number of APIs, relative to no medication use, but only if at least one of the APIs belonged to medication classes that were used less commonly.Furthermore, higher numbers of APIs were related to the most prominent increase in mortality risk when the study participants were below the age of 60 and to a smaller increase when they were between ages of 60 and 70 years, whereas above the age of 70 the number of APIs no longer showed a clear association with mortality risk.Overall, our data indicate that API count may be a useful predictor for mortality risk, especially below the age of 70 years.Reviews and meta-analyses of epidemiologic studies showed an average relative mortality risk of about 1.3 for individuals falling into a "polypharmacy" category, mostly defined by use of ≥ 5 medications, relative to individuals using a smaller number of medications 6,8,9 .Based on a smaller subset of studies, meta-analyses further showed an all-cause mortality HR of 1.08 per single additional medication used, and showed a pooled adjusted risk ratio of 1.24 (95% CI 1.10-1.39)for studies using a threshold of 1-4 medications, 1.31 (1.17-1.47)for 5 medications, 1.59 (1.36-1.87)for 6-9 medications, and 1.96 (1.42-2.71)for 10 or more medications 6,7 .The studies included in these meta-analyses reported associations of variable strength between number of medications used, polypharmacy and mortality risk, and were also highly diverse with regard to overall study size, types of study population (e.g., community-dwelling individuals within variable age ranges; people living in care institutions; geriatric patients; hospital discharge patients), and different definitions for individuals classified as being exposed to polypharmacy or to a comparison (reference) group.In addition, studies varied widely regarding adjustments made for confounding by indication or for confounding by general health-related risk factors: some of studies included adjustments for selected comorbidities or for more comprehensive multi-morbidity scales such as Charlson's Comorbidity Index.Other studies used more novel methods such as propensity score matching to assess whether polypharmacy itself had causal effects on mortality risk independently of the underlying health conditions.The heterogeneity in the analytical approaches used may reflect differences in the original study objectives, either aiming to examine medication use as a potential cause of death, or rather focusing on medications prescribed as a substitute indicator for health conditions in the prediction of mortality risk.In our analyses, we chose to calculate HRs without any adjustment for self-reported co-morbidities.We believe that medication data may be used merely as an instrument to predict mortality risk and life expectancy, as a substitute indicator for underlying morbidities that are the more likely risk determinants.
A general issue with the studies modeling adverse health effects of polypharmacy as a monotonous function of a mere count of different APIs used, is the implicit assumption that each additional API, irrespective of its type, would increase the risk of the adverse effect to an equal extent.Many studies may have used this assumption to allow any form of statistical modeling at all, as often these studies were too small to allow modeling in relation to more specific categories of medication exposures.However, our observation that mortality risk was associated more strongly with APIs that were used less commonly, as compared to commonly used APIs, indicates that this assumption represents an over-simplification that may lead to suboptimal modeling of adverse health effects.Recent analyses within very large-scale, national drug prescription databases in Sweden have provided further confirmation for this, showing relatively strong risk associations when focusing on more specific medication types at the subgroup (3rd or 4th) coding level of the ATC classification, but less clear-cut associations for other medications within a same class at the more aggregated, 2nd ATC coding level 11,12 .In these Swedish studies, very strong associations with mortality risk were noted, for example, for serotonin antagonists (N06AB) and for propulsive drugs (A03F), frequently prescribed for nausea and vomiting and likely indicators for malignancy and associated treatment, or sub-types of diuretics (C03CA, C03DA) prescribed for severe conditions such as heart and kidney failure, whereas many other drugs in the "N", "A" or "C" categories of the ATC system showed  www.nature.com/scientificreports/no significant association with risk.These examples underlined the importance of information carried in the 3rd and 4th levels of the ATC code (specifically: the pharmacological subgroup, over the information carried by the main anatomical group) when using medication data for the purposes of predicting mortality risk 11,12 .
Unfortunately, in spite of including a fairly large number of (> 4500) mortality events out study was far too small to allow analyses on mortality risk in relation to a more detailed and more specific classification of APIs on the 3rd or 4th level of the ATC classification, as was done in the Swedish national database studies.Nonetheless, our findings do show that, to a large extent, associations between medication use and increased risk of death may be driven by less commonly used drugs that may to indicate more serious underlying medical conditions.Interestingly, however, our data also indicate that especially for individuals using at least one less common medication type, risk of death is also associated with total API count.The latter suggests that also the more common medication types (e.g.analeptics, psycholeptics) may gain predictive power for mortality risk for individuals treated for a more serious morbidity condition.
Compared to previous studies, a unique feature of our investigation is that it was based on regular, prospective data collections over a 20-year period, for a community-dwelling population of women and men who were initially between 35 and 65 years of age, applying a time-dependent modelling approach to examine the prospective relationships between (changes in) medication use and mortality risk.This longitudinal approach, combined with fairly large number of mortality outcomes over time, allowed us to examine risk associations within different age windows, as participants in the EPIC-Heidelberg grew older.Medication use was defined by a numerical threshold depending on the number of APIs contained in each medication, as a more precise definition 19 .Our data also allowed careful adjustments for possible confounding by relevant lifestyle factors.An interesting finding from our analyses is that with advancing age the association of mortality risk with number of APIs used progressively appeared to flatten, a phenomenon that, to our knowledge, has not been given much attention in previous studies on multiple medication use, polypharmacy scores and mortality risk.While polypharmacy, defined as the use of 5 or more APIs, reached a high prevalence above the age of 70 years, we also observed that 80% of medications within this age group were concentrated in only 12 frequently used API classes.Most of these medications were for APIs used as preventive medications, e.g. to avoid worsening of highly prevalent conditions related to the cardiovascular system (e.g.hypertension) and blood coagulation, or used as pain killers to combat symptoms related to muscular-skeletal problems, which are mostly low-risk conditions.Other studies have also shown that among elderly individuals with high prevalence of medication use, most medications belonged to a very restricted number of ATC classes [22][23][24][25][26] .Furthermore, it has been reported that among elderly individuals (residents of nursing facilities) the association of polypharmacy with mortality risk may be modified by other health conditions such as frailty, and that risk of death was increased mostly by frailty and disability, whereas (hyper-)polypharmacy was associated with higher risk of death only among non-frail study participants 27 .
A minor limitation of our investigation is the proportion of missing medication information across followups, which adds up to 23% of at least one time missing medication use, although only a small proportion of study participants had missing information on three or more follow-up occasions (5%).Theoretically, a selective drop-out of study participants could have biased our hazard ratio estimates for mortality, if absence of response to a follow-up questionnaire was non-random and in itself informative about an individual's health status ("informative missings").However, information from follow-up questionnaires was carried forward only for a limited amount of time, which very much should mitigate any risk for such bias.Also, to further explore informative missingness as a potential source of bias, we additionally applied more complex approaches for the joint modelling of medication exposures and mortality risk 28 .The results from this more complex modelling (results not shown) were practically identical to those from the simpler time-variable Cox modeling presented in the present manuscript.
In summary, our longitudinal analyses show a complex quantitative relationship between multiple medication-use and risk of all-cause mortality in a general population cohort of middle-aged and older individuals.Our data indicate that classifying individuals into risk categories based on the mere number of medications used, as with the "polypharmacy" concept, represents an over-simplification that may lead to sub-optimal modeling of adverse health effects.More accurate risk modeling would require accounting for a higher level of detail and specificity for medication types described for severe vs. less severe health conditions.Furthermore our data show that with advancing age, as multiple medication use becomes increasingly prevalent, the prediction of mortality risk becomes increasingly complex and may be less directly related to a mere number of APIs used.This phenomenon, and the specific relationships of mortality with low-vs.high-risk medication types, should be explored further in context of very large-scale studies, e.g. using national drug prescription databases 11-13,23-26 . https://doi.org/10.1038/s41598-023-50487-5

Table 1 .
Baseline characteristics of the EPIC-Heidelberg cohort (N = 25,540).*Number of active pharmaceutical ingredients taken.

Table 2 .
Hazard ratios (and 95% CI) for API count and mortality across categories, for individuals using all API, only common API, or at least one uncommon API.Crude model-adjusted for age and sex.Full modeladjusted for age, sex, BMI, smoking history (average daily cigarettes, duration, and time since quitting), physical activity, highest level of schooling.Death count is the number of individual deaths while in the specific API category.Individuals can move across API categories throughout follow-up.a Active pharmaceutical ingredient.b Reference category is no API intake.Hazard ratios (with 95% confidence intervals) for API count and mortality across age categories, for individuals using at least one uncommon API compared to no API intake.