Temporal validation of a nationwide target trial emulation of cholinesterase inhibitors and memantine add-on in the Brazilian Unified Health System (SUS)

This project studies the comparative effectiveness of the three cholinesterase inhibitors used to treat Alzheimer’s disease in the Brazilian Unified Health System (SUS): donepezil, galantamine and rivastigmine. The SUS dispenses all three free of charge to people with mild to moderate disease. No national evidence has been available to guide the choice between them, and few randomized trials have compared them head to head or followed patients for more than one year. Purpose. The study emulates a pragmatic three-arm target trial using the outpatient medication dispensing records (APAC) of the Outpatient Information System (SIA/SUS), which the Department of Informatics of SUS (DATASUS) publishes monthly for all 27 Brazilian federative units. The outcome is memantine add-on. Under the national clinical protocol, memantine is added to a cholinesterase inhibitor when the disease reaches the moderate stage, so time to add-on is a pragmatic marker of clinical progression that is recorded for every treated patient. Design and methods. New users with a main diagnosis of Alzheimer’s disease (ICD-10 G30 or F00) who started a single cholinesterase inhibitor between October 2019 and December 2025, after at least 12 months without anti-dementia treatment, were followed from the first dispensing (time zero). Confounding by recorded baseline characteristics was addressed with stabilized inverse probability of treatment weights from a multinomial propensity model. Loss to follow-up was addressed with inverse probability of censoring weights. Confidence intervals were obtained from 200 bootstrap replicates. A pilot study in four states (2018 to 2023) tested file integrity, identifier stability, clinical coding and cohort feasibility against prespecified go/no-go criteria before the national analysis. The design changes made after the pilot are documented with dates. The study is reported following the STROBE statement and the TARGET guideline. Main findings of the discovery cohort. The cohort included 142,758 new users: 74,165 started donepezil, 30,691 galantamine and 37,902 rivastigmine. The 24-month weighted cumulative incidence of memantine add-on was 9.7% with donepezil, 9.0% with galantamine and 5.6% with rivastigmine. The risk difference for rivastigmine versus donepezil was -4.1 percentage points (95% CI -4.5 to -3.7). The difference was larger with the transdermal patch than with oral rivastigmine, and it increased with age. It persisted with a composite outcome that counts exit from the dispensing records as an event. Deaths were almost never recorded in the dispensing data, and disease severity is not recorded. The finding may therefore reflect residual confounding or prescribing patterns rather than slower progression. Expected outcomes and next steps. Because a difference of this size was not expected from trial evidence, the project includes a preregistered temporal validation. The confirmatory analysis will use a new, non-overlapping cohort of patients who start treatment between January 2026 and December 2027. It follows a protocol registered before any SIA/SUS file after December 2025 was downloaded. It tests whether rivastigmine reduces the 12-month risk of memantine add-on relative to donepezil, and whether the effect varies with age and formulation. Results will be reported whatever their direction. Further work will examine alternative explanations: 1. an instrumental variable based on the prescribing preference of the dispensing establishment; 2. comparisons within the same establishment; 3. linkage with the Mortality Information System, if authorized. The project also assesses biological plausibility in human single-nucleus transcriptomic data. Contents. The project contains the preprint and a complete replication package: 1. code for download, pseudonymization, data quality checks, cohort construction, weighting, bootstrap, figures, tables and the manuscript; 2. aggregate results and download manifests; 3. the dated log of design decisions; 4. the pilot study materials; 5. the preregistered protocol for the temporal validation. Data protection. No patient-level data are shared. The health card number in the public files is replaced at download by a keyed HMAC-SHA256 pseudonym, and counts of 1 to 4 are shown as “<5”. Research with public-domain data of this kind does not require ethics committee review under Resolution 510/2016 of the Brazilian National Health Council.