Drugs for Dementia and Excess of Hospitalization: A Longitudinal French Study

Data

The databases used in this study were the Echantillon gé né raliste des bé né ficiaires (EGB) and the Programme de mé dicalisation des systè mes d’information (PMSI). The EGB is a sample of a larger database, the French national health insurance information system (SNIIRAM), which is a health administrative database providing exhaustive information on all reimbursed ambulatory care of people affiliated with any of the main French health insurance funds. Every individual in the SNIIRAM is identified by his health insurance number (NIR), which includes a control key number that varies from 1 to 97. Individuals included in the EGB are those for whom the control key number of the NIR is equal to a given value, which is confidential. Thus, the EGB represents a subsample comprising 1/97th of the SNIIRAM. The EGB constitutes a representative sample of the population covered by the French health insurance in terms of age, gender, and geographical location, in which more than 700,000 individuals are included, even if they do not consume any health care service [20]. The primary strength of the EGB is that the data are stored for several years and have been available for research since 2006 for the general scheme, further called ‘Ré gime Gé né ral’, and since 2011 for the other French social security schemes. Patients covered by the Ré gime Gé né ral account for 82% of patients in the EGB. The main limitation of the EGB is that clinical information is not available. The PMSI is a medico-administrative hospital discharge database set up to evaluate costs of hospital stays according to different groups of patients (Diagnosis Related Groups, DRG) differing by their medical procedures, duration of stays, comorbidities. The PMSI provides exhaustive information on hospital care in France such as diagnoses (coded by the physician using the International Classification of Diseases, 10th revision (ICD-10)), underlying comorbidities and potential complications, length of stay, etc. Each patient’s stay is classified by DRG according to the information documented by the physician. Costs of hospital stays have been available for research since 2011. PMSI data are provided to the French public health insurance fund for linkage with the SNIIRAM, through the national identification number (NIR). However, the NIR is not accessible in the SNIIRAM in order to preserve the identity of patients. It is made anonymous by a complex two-level procedure, which is completely irreversible. EGB and PMSI data are available in the same interface.

Study population

The study population was an unbalanced panel composed of patients suffering from dementia aged 65 and over and affiliated with the Ré gime Gé né ral between 2007 and 2014. Patients were included from the diagnosis of dementia until their death or until 2014, if they were still alive. For instance, a person born in 1942 was included in 2007 if dementia was diagnosed in 2007 or earlier, or was included from dementia diagnosis if made between 2007 and 2014.

Patients suffering from dementia were identified in three ways. The first way was the presence in the EGB of dementia declared as a long-term disease (LTD). Dementia is one of the 30 LTD listed by the French health insurance. LTD corresponds to diseases for which the severity and/or the chronicity require a long-term and particularly costly treatment. For patients with LTD, healthcare costs are entirely covered by the public health insurance. A limitation is that the type of dementia could not be distinguished. The second way was to identify three distinct deliveries of ChEIs and/or memantine during the study period for a same patient. The third way was to spot in the PMSI a patient with a diagnosis of dementia (as main diagnosis or underlying comorbidity) during a hospitalization [32]. Using these three criteria, a patient suffering from dementia was included in the study either 1) from the declaration of dementia as a LTD, 2) the first three distinct deliveries of ChEIs or memantine, or 3) the first hospitalization with a diagnosis of dementia.

Assumption to be tested

The assumption being tested in this study is whether or not patients treated for dementia are more often hospitalized than non-treated patients, especially with regard to hospitalizations where the reason could be one of the adverse effects of treatment.

Variables

Once the study population was selected, several endpoints were tested (all-cause hospitalization, hospitalization for cardiac, gastrointestinal, and psychiatric adverse effects) for various subgroups (pooled across all treatments for dementia, and then stratified by substance) according to the following steps:

All-cause hospitalization for all ChEIs and memantine,

We considered the following reasons for hospitalization due to adverse effects [16, 33]:

Cardiac: bradycardia, rhythm disorders, cardiac conduction disorders, hypotension, cardiac arrest, syncope, myocardial infarction.

In all the regression models, the main independent variable was the reimbursement of a treatment for dementia, i.e., ChEIs (including donepezil, galantamine, and rivastigmine) or memantine at least once in the previous quarter of hospitalization.

The adjustment variables were age, gender, number of LTDs (other than dementia), and the “Accessibilité potentielle localisé e” (APL). APL indicator measures the supply of and demand for general practice services by taking into account practitioners’ volume of activity on the one hand, and service use rates differentiated by population age structure on the other. A decreasing function of services use with distance is also used to weight spatial accessibility according to the remoteness of the population to health care supply [34]. The APL was calculated in 2010 and was expressed in terms of full-time equivalent per 100,000 inhabitants [35].

Statistical analysis

Because the analysis of hospitalization in relation to exposure to ChEIs or memantine needed to take place in a close timeframe, each year of observation between 2007 and 2014 was cut into quarters, and patient-quarter was used as time unit.

Binary logit models for longitudinal data were estimated as follows: logit ( p [ y i , t ] = 1 ) = β 0 + β 1 treatment i , t - 1 + β 2 age i , t + β 3 age i , t 2 + β 4 gender i + β 5 other LTD i , t - 1 + β 6 APL where p[y_i,t] denotes the probability to be hospitalized at least once for a reason y (hospitalization for all causes or for adverse effects), for an individual i at quarter t; treatment_i,t - 1 is the reimbursement of ChEIs or memantine (yes/no) at least once for an individual i at quarter t-1; age_i,t is the age of an individual i at quarter t; age²_i,t is the square of age of an individual i at quarter t, gender _i is the gender of an individual i; and other LTD_,t - 1 is the number of other LTDs for an individual i at quarter t-1.

Age was included as a continuous variable together with a squared term in order to account for a potential non-linear impact of age on the probability of hospitalization. Indeed, we suspected that age would increase the probability of hospitalization (β₂ >0) but at a decreasing rate (β₃ <0): oldest patients, reaching these advanced ages precisely because they are in better health, would probably be less at risk of hospitalization. The number of associated LTD was included as a linear variable and we suspected an increase in the number of LTDs could increase the probability of hospitalization (β₅ >0). The mean APL in 2010 was 71 full-time equivalents per 100,000 inhabitants. The APL indicator was categorized in three classes: 0–59 (low), 60–79 (medium), and over 80 (high) [35].

The longitudinal binary logit models used were implemented using the generalized estimating equations (GEE) technique by specifying a logit function and auto-regressive correlation structure to account for time-dependence among repeated observations. The GEE approach models the marginal distributions of dependent variables and provides consistent estimators of coefficients and robust standard errors in binary models with longitudinal data [36]. The auto-regressive correlation structure allowed observations to be correlated for the same individual over time, with the highest estimated correlations for close time points (e.g., first and second quarter in 2011) and lower estimated correlations for longer gaps between time points (e.g., first quarter in 2008 and third quarter in 2014). In a supplementary analysis, we tested the robustness of the models using a random intercept binary logit model in order to account for unobserved individual heterogeneity.

The software used for the analysis was SASGUIDE 7.1.

Cost analysis

The additional cost of hospitalization for adverse effects related to treatments, adjusted for the structure of the EGB population (age, gender, number of LTD, APL), was calculated as the difference between the expectation of cost for treated and non-treated patients as follows: E ( C | T ) - E ( C | NT ) = P ˆ hosp | T × C hosp + ( 1 - P ˆ hosp | T ) × C nhosp - [ P ˆ hosp | NT × C hosp + ( 1 - P ˆ hosp | NT × ) × C nhosp ] where E(C|T) is the expectation of cost for treated patients, E(C|NT) the expectation of cost for non-treated patients, P ˆ hosp | T the probability to be hospitalized for a treated patient, P ˆ hosp | NT the probability to be hospitalized for a non-treated patient, C_hosp the total cost of hospitalizations for adverse effects, and C_nhosp the total cost of non-hospitalizations for adverse effects.

As C_nhosp was null, the equation simplifies as follow: E ( C | T ) - E ( C | NT ) = P ˆ hosp | T × C hosp - ( P ˆ hosp | NT × C hosp ) = C hosp × ( P ˆ hosp | T - P ˆ hosp | NT )

To estimate P ˆ hosp | T and P ˆ hosp | NT , the estimate of the binary logit models for hospitalization for adverse effects was used as follows: P ˆ hosp | T = exp ( T ) 1 + exp ( T ) with T = ∑ k = 1 K γ k X k + β 1 treatment i , t - 1 where x_k denote each adjustment covariate (age, gender, number of other LTD, APL) P ˆ hosp | NT = exp ( s ) 1 + exp ( s ) with S = ∑ k = 1 K γ k X k

C_hosp, the total cost of hospitalization for adverse effects, was calculated by adding up the costs of hospitalization for adverse effects between 2011 and 2014 for treated or non-treated patients (cost of hospitalization were not available before 2011). Costs were expressed in euros 2014. Once the additional cost of hospitalization for adverse effects related to treatments was obtained in our sample, this result was extrapolated to the entire EGB dividing by 0.82 (proportion of individuals aged 65 and over covered by the Ré gime Gé né ral in the EGB) then extrapolated to the French general population by multiplying by 97, as the EGB is a subsample comprising 1/97th of the population [20].

Ethics

Access to data is subject to prior training and authorization and has received approval from the independent data protection administrative authority (Commission Nationale Informatique et Libertés).

Descriptive statistics

The descriptive statistics of the study population are presented in Table 1. Over the period 2007–2014, 7,668 patients were followed. Among them, 5,809 had a dementia declared to the French health insurance, 5,059 received treatment at least three times for dementia, and 489 were hospitalized with a diagnosis of dementia. In total, 111,133 observations were generated.

Statistical analysis

Results of costs analysis

The probabilities to be hospitalized for cardiac events were P ˆ hosp | T = 0.43 % and P ˆ hosp | NT = 0.32 % for treated and non-treated patients, respectively. The probabilities to be hospitalized for gastrointestinal events were P ˆ hosp | T = 0.15 % and P ˆ hosp | NT = 0.10 % for treated and non-treated patients, respectively. The probabilities to be hospitalized for psychiatric events were P ˆ hosp | T = 0.31 % for treated patients and P ˆ hosp | NT = 0.40 % for non-treated patients, respectively. The total cost of hospitalization for cardiac events, for gastrointestinal events, and for psychiatric events between 2011 and 2014 were € 2.1 million, 0.3 million, and 0.7 million, respectively. Between 2011 and 2014, the additional cost of hospitalization for cardiac and gastrointestinal adverse effects, and for psychiatric events related to treatment was estimated at € 2,420, € 140, and € -710, respectively. Extrapolated to the entire population, this cost was estimated at € 220,000 (subtracting € 84,000 of hospital costs for psychiatric events avoided), which is € 55,000 per year.

Strengths and limitations of the study

Our study relied on observations from a medico-administrative database. Its size and its longitudinal nature enabled us to observe non-frequent events (such as hospitalization for gastrointestinal events under rivastigmine). Its national representativeness enabled the findings to be generalized to the French population. However, the study has some limitations. First, because clinical data were not available to identify patients with dementia, we selected patients suffering from dementia when they were declared as such to the French national health insurance, and/or when they were prescribed drugs for dementia, and/or when there were hospitalized with a diagnosis of dementia. This could have led to selection bias by selecting only the most severe patients. Indeed, non-treated and non-hospitalized patients without dementia declared to the national health insurance have not been included in our study. However, we believe that these patients were not numerous. Second, it would have been interesting to adjust the results on the severity of dementia. A proxy could have been the date of dementia declaration as a LTD. Unfortunately, the date of declaration usually does not correspond to the starting date of dementia thus the information would have been biased. While some physicians may declare dementia only when treatments for dementia are prescribed, others may declare dementia immediately following the onset of symptoms. Furthermore, more than 25% of the study population had no dementia declared. Thirdly, although Alzheimer’s disease is the most common cause of dementia [53], dementia types could not be distinguished. This was not an issue for Lewy bodies dementia which can be treated by the same drugs than Alzheimer’s disease or for Parkinson’s dementia which can be treated by rivastigmine. This could be an issue for patients with vascular dementia who were at the same time more likely to be hospitalized for cardiac events given their pejorative cardio-vascular health status, and over-represented in the non-treated group, and, therefore, potentially confusing. We believe we controlled this phenomenon by adjusting for the number of LTD (including cardio-vascular diseases) associated with dementia. In any case, if not entirely controlled, this would mean our findings could be under-estimated. Finally, duration and severity of dementia were also not available. Given that severity of dementia is a determinant of hospitalization and that memantine is actually recommended in severe patients [51, 54], this could have led to overestimate the impact of memantine on excess of hospitalization.

Conclusions

The prescription of drugs for dementia, more specifically of rivastigmine, increases the risk of hospitalization via their cardiac and gastrointestinal adverse effects among a population already highly exposed to the risks of overuse. Instead, the objective should be to reduce the number of hospitalizations in these fragile populations. These prescriptions also lead to additional health care expenditures. Even if more studies are needed to confirm the impact on hospitalization of gastrointestinal adverse effects, these results may contribute to feed in the current debates on the relevance of these pharmacological treatments in the management of dementia and encourage considering with caution the balance between benefits and harms before prescription.