In the present study, we undertook two different observational cohort studies based on nationwide population-based databases in Denmark and Germany, to analyze the association between the prescription of TNFi and incident dementia in older adults with rheumatological disorders. We observed a trend towards a lower risk of dementia among patients treated with TNFi drugs versus patients treated with other anti-inflammatory drugs; however, the association did not reach statistical significance.

As sensitivity analyses, we tested five other treatment groups (non-TNF biologics, corticosteroids, low-potency immunosuppressants, intermediate-potency immunosuppressants, and high-potency immunosuppressants), which are commonly used in rheumatological systemic treatment and have anti-inflammatory properties through other mechanisms than targeting TNF-α. Only the group with low-potency immunosuppressants in DANBIO showed a statistically significant association (HR = 0.84; 95% CI 0.72, 0.99). None of the other analyses reached statistical significance, which is similar to other epidemiological studies and clinical trials, although it is an area with conflicting results [24,25,26,27].

Of note, we performed further sensitivity analyses, where patients treated with one of the six treatment groups (including TNFi) were compared with a ‘mixed control group’ that consisted of individuals aged ≥ 65 years taking any other prescription drug or no drugs. All analyses reached a significant association with a decreased risk of dementia. Thus, we were able to reproduce the results of the study by Chou et al. [11]. This small nested case-control study was based on health claims data from 2000 to 2007 and showed that only etanercept was associated with a lower risk of AD in patients with RA compared with other therapies for RA with a ‘mixed control group’. One possible argument for the increased risk of dementia for patients without prescriptions of the analyzed drugs may be the healthy user effect. This analysis shows the importance of a suitable control group for the interpretation of study results.

Relatively few epidemiological studies have so far investigated treatment with TNFi and incident dementia [10,11,12,13,14,15,16]. The available data suggest that TNFi may be promising candidates for the treatment of dementia, but there are methodological limitations to the prior studies in terms of the observation period and study design [11,12,13,14].

The TNF pathway appears to be specifically implicated in AD. Gene prioritization in the new loci identified 31 genes that were suggestive of new genetically associated processes, including the TNF-α pathway through the linear ubiquitin chain assembly complex [28]. Regarding the specific question of TNFi and central nervous system penetration, we can imagine at least two different pathophysiological scenarios. First, AD has been specifically associated with the breakdown of the blood–brain barrier early in its course (see [29] for further on this topic), which promotes the entry of a broad array of moieties, including AD-specific drugs. Of note, drugs such as etanercept have nearly identical molecular weights to AD-targeting agents such as aducanumab. Second, it is possible that TNFi act strictly in the periphery, but this results in systemic changes that secondarily affect dementia, perhaps by affecting other small-molecule members of the inflammatory cascade.

One strength of our study is the cohort design in two population-based databases. Even though we had up to 20 and 9 years of follow-up in DANBIO and AOK PLUS, respectively, the median follow-up for both cohorts approximated 5 years. However, as dementia possibly has a long prodromal phase with neuroinflammatory and inflammatory pathology many years before diagnosis [30], it seems likely that the maximal benefit from the anti-inflammatory properties of TNFi might require a longer follow-up and earlier treatment. Zheng et al. also had a long follow-up of 14 years (median) among patients with RA and found no difference in the effect over time [15]. Yet, more studies are warranted with longer follow-ups.

Drug combinations are commonly used in rheumatological treatment and a limitation in our study is that it is difficult to disentangle the complexity of anti-inflammatory drug treatment. As a result, the patients in our study were often treated with multiple drugs of interest at different timepoints through the follow-up. According to current RA treatment guidelines, a combination of traditional disease-modifying antirheumatic drugs (high-potency, intermediate-potency, and low-potency immunosuppressants) or TNFi/methotrexate (high-potency immunosuppressants) combinations are suggested for patients with moderate or high disease severity [22]. We tested multiple anti-inflammatory drug classes, which could partially be an indication of different degrees of the severity of rheumatological disease. Confounding by indication is common in observational studies and it is difficult to separate whether patients receiving TNFi treatment are different from patients taking other anti-inflammatory treatments, both in disease severity and underlying pathophysiology of the rheumatological disease, especially given that these are highly heterogeneous and complex diseases.

An additional bias that could be part of our study is the healthy user effect. This bias may derive from both patient and clinician factors, both of which are possible in our analyses. For example, medication adherence has been associated with better health in several studies [31, 32] . Similarly, clinicians may elect only to prescribe medications that require long-term adherence and monitoring to patients perceived to have a favorable long-term prognosis outside of their autoimmune conditions, leading to a lower observed risk of future dementia. Our results using medication non-users tend to validate this concern.

Furthermore, a limitation in both cohort studies was that the definition of drug use was based on prescriptions, and it was not possible for us to assess compliance with the prescribed drugs.

However, we were fortunately able to create a reasonably homogenous dataset that reduced confounding by only including patients with rheumatic diseases and systemic therapies. As a result, within this patient cohort, even a single prescription code, for example, is likely to be classified correctly in nearly every instance.

In an intention-to-treat framework, we assumed that patients with a drug prescription started at the time of the first prescription and continued until they were censored from the study or end of study. In addition, we did not investigate details on drug usage duration, dosage, and, as mentioned before, drug combinations. Patients normally do not get biologics as first-line therapy for rheumatic diseases. A new-user design could therefore only include patients with second-line therapies, which would exclude a large proportion of the patient group with systemic medications. Because of different starting dates of the TNFi therapy before dementia, no cumulative dose of TNFi was counted in the analyses.

Moreover, our primary outcome of interest was all-cause dementia because AD frequently coexists with other dementia types such as vascular dementia. Detailed information on dementia subtypes is not available in disease registers. Our dementia definition was solely based on the first documented ICD code, which probably leads to an underdiagnosis and a misdiagnosis in both DANBIO and AOK PLUS, especially for the early stages of the disease. However, previous evidence [33] using a single billing claim for dementia found sensitivity and specificity values of 85–89%, suggesting a high degree of validity.

In addition, we did not have sufficient numbers of cases of diseases other than RA to compare the effects across rheumatological disorders. In AOK PLUS, 90% of patients belong to the RA group, 8% to the SA, and 4% to the PsA. About 0.9% had RA and SA and 0.6% had RA and PsA. Similarly, in DANBIO, the distribution of diagnoses was as follows: 92% RA, 4% SA, and 4% PsA. About 0.6% had RA and SA and 0.8% had RA and PsA. Finally, we were not able to control for all known risk factors, such as apolipoprotein ε4 status, smoking, alcohol consumption, and other lifestyle factors, because this information was lacking in both databases.