Long‐term risk of dementia following hospitalization due to physical diseases: A multicohort study

Abstract Introduction Conventional risk factors targeted by prevention (e.g., low education, smoking, and obesity) are associated with a 1.2‐ to 2‐fold increased risk of dementia. It is unclear whether having a physical disease is an equally important risk factor for dementia. Methods In this exploratory multicohort study of 283,414 community‐dwelling participants, we examined 22 common hospital‐treated physical diseases as risk factors for dementia. Results During a median follow‐up of 19 years, a total of 3416 participants developed dementia. Those who had erysipelas (hazard ratio = 1.82; 95% confidence interval = 1.53 to 2.17), hypothyroidism (1.94; 1.59 to 2.38), myocardial infarction (1.41; 1.20 to 1.64), ischemic heart disease (1.32; 1.18 to 1.49), cerebral infarction (2.44; 2.14 to 2.77), duodenal ulcers (1.88; 1.42 to 2.49), gastritis and duodenitis (1.82; 1.46 to 2.27), or osteoporosis (2.38; 1.75 to 3.23) were at a significantly increased risk of dementia. These associations were not explained by conventional risk factors or reverse causation. Discussion In addition to conventional risk factors, several physical diseases may increase the long‐term risk of dementia.


INTRODUCTION
Because no treatment to stop or delay the progression of dementia is available, prevention remains the primary means to combat the disease. 1,2 The National Academies of Sciences, Engineering, and Medicine recommend three targets for dementia prevention: cognitive training, management of hypertension, and physical activity. 3 The 2017 Lancet Commission, 1 and more recently, the World Health Organization 4 have highlighted further potentially modifiable risk factors, including low education, obesity, hearing loss, smoking, depression, social isolation, and diabetes. The associations between these risk factors and dementia are, however, relatively modest, with metaanalyses reporting relative risk estimates of two or below. 1,[5][6][7][8][9][10][11][12] Many people with dementia are affected by multiple other physical diseases 13,14 that may also increase the risk of dementia, although they are not currently acknowledged as targets for dementia prevention.
These include cardiovascular disease and some autoimmune, inflammatory, and infectious diseases. [15][16][17][18][19] Genetic studies, supporting the role of these diseases in the development of dementia, have found shared genetic variants for lipid metabolism, the immune system, and Alzheimer's disease, a common form of dementia. 20,21 To date, however, few studies have systematically compared the magnitude of the dementia risk associated with conventional risk factors and physical diseases.
To address this limitation, we examined the association between 22 potentially dementia-related physical diseases 22 and incident all-cause dementia before and after adjustment for conventional risk factors.
Using a data-driven approach, our aim was to generate new hypotheses for research on the prevention and etiology of dementia. To minimize ascertainment and reverse causation biases, we excluded the first 10 years of follow-up in the sensitivity analyses. 10

Study population
This analysis is based on four prospective cohort studies linked to national health registries: the Finnish Public Sector study (FPS), the Health and Social Support study (HeSSup), and the Still Working study (STW) in Finland, and the Whitehall II study (WHII) in the UK. Of the 351,331 participants eligible for this study, we selected adults aged 18 years or older, who were successfully linked to hospitalization reg-

Diagnoses of hospital-treated diseases
We linked participants to the records of national health registers and retrieved both primary and secondary diagnoses from inpatient hospi-

HIGHLIGHTS
• tal discharge information using the International Classification of Diseases, 10th Revision (ICD-10). The diagnosis codes from the 8th and 9th revisions (ICD-8 and ICD-9) were converted into the corresponding ICD-10 codes (Table A.1 in the Appendix). To avoid false-positive findings and publication bias, we systematically considered 32 physical diseases that were associated with an elevated 5-year incidence of at least one type of dementia (or were on a disease trajectory later leading to dementia) in a hypothesis-free Danish total population study with rigorous control for multiple testing. 22 We excluded diseases in ICD-10 Chapters V (Mental and behavioral disorders) and VI (Diseases of the nervous system) because their relation to dementia has already been studied extensively. 9,23,24 We found sufficient case numbers in our data for 25 of these physical diseases (i.e., a minimum of three incident dementia cases that occurred more than 10 years after the diagnosis of the disease). Current prevention guidelines 1,4 acknowledge three of these diseases (type 1 and 2 diabetes and hypertension) as dementia risk factors, so we examined 22 other physical diseases as potential risk factors for dementia ( Figure 1).

Covariates
We included the following covariates: 1

Ascertainment of incident dementia
We focused on all-cause dementia because electronic health records may not be a reliable source of data to ascertain specific dementia sub-

Statistical analysis
We computed hazard ratios for the associations of hospital-treated physical diseases (index diseases) with dementia in separate Cox proportional hazards models. We used a one-stage approach for a pooled analysis of individual-level data from four cohort studies and took the within-study clustering of participants into account using cohort-specific baseline hazards and cohort-specific adjustment terms for covariates. 25 This means that individual-level data from the four cohorts were analyzed in a single Cox model that was stratified by cohort. The model produced a single effect estimate for the index disease, but the effect estimates for adjusting variables (sex in all models and conventional dementia risk factors in further adjusted models) were allowed to vary from cohort to cohort. All Cox models were adjusted for age using age as the time scale.
Follow-up for incident dementia started at study entry and continued until dementia diagnosis, death, or end of follow-up, whichever came first. We modeled exposure to the index diseases using timedependent analysis whenever hospitalization from the index disease occurred after study entry (incident index disease). 26 If hospitalization occurred before study entry (prevalent index disease), the participant contributed to the follow-up as exposed from study entry. We also performed sensitivity analyses restricted to those free of the index disease at study entry and after excluding dementia cases that lacked information on the type of dementia. We examined the proportional hazards assumption using scaled Schoenfeld residuals (Methods A.2 and Figure A.2 in the Appendix). We analyzed early and late-onset dementia as separate outcomes by splitting follow-up at age 65 (those whose dementia onset was before age 65 vs those whose dementia onset was at or after age 65).
To reduce the risk of ascertainment bias and reverse causation, we repeated the analysis after excluding incident dementia cases that occurred during the first 10 years after initial hospitalization for the index disease (these biases mainly affect cases of dementia diagnosed early in the follow-up). 10 We examined whether the index diseases were associated with dementia independently of each other by entering them into the same Cox model. Two diseases reflected coronary heart disease: myocardial infarction and chronic ischemic heart disease. To avoid overadjustment, these two diseases were not adjusted for each other.
We used the Fine-Gray competing-risks model to compute subhazard ratios with death without dementia as the competing outcome event. The Fine-Gray model was specified analogously with the Cox models. Age was the timescale, and the effect estimate for sex (the adjusting variable) was allowed to vary from cohort to cohort. We also adjusted the models for cohort and used robust standard errors that were clustered for cohort.
To compare the relative dementia risk related to physical diseases versus conventional potentially modifiable dementia risk factors (low education/socioeconomic status, hypertension, obesity, smoking, depression, physical inactivity, marital status, and diabetes), we computed the hazard ratios for these risk factors using the same modeling as for the index diseases. Because blood pressure and weight declines in preclinical dementia, we analyzed hypertension and obesity at middle age, excluding those aged 65 or older at baseline. 29,30 Analogously with the models for physical diseases, individual-level data from the cohorts were analyzed in a single Cox model that produced a single effect estimate for the risk factor of interest and was stratified by cohort. Age was the timescale and the models were adjusted for sex, which was allowed to have different effect estimates for different cohorts.
To assess the effect of conventional dementia risk factors on the associations between physical diseases and dementia, we adjusted the Cox models for these factors, sex, and the statistically significant inter- Within-study clustering of participants was taken into account using cohort-specific baseline hazards and cohort-specific adjustment terms for sex. *Significant association in the main analysis after Bonferroni correction for 22 tests (P < .0023). †Significant association after Bonferroni correction in the main analysis and significant association (P < .05) in the sensitivity analysis excluding first 10 years of dementia follow-up. Within-study clustering of participants has been taken into account using cohort-specific baseline hazards and cohort-specific adjustment terms for sex. *Significant association in the main analysis after Bonferroni correction for 22 tests (P < .0023). † Significant association after Bonferroni correction in the main analysis and significant association (P < .05) in the sensitivity analysis excluding first 10 years of dementia follow-up.
actions between the adjusting variables (indicated by a likelihood-ratio test with P < .05) in a subcohort with relevant data available (Table A.2 in the Appendix). One cohort (STW) was excluded from these analyses because no data on body mass index were available. As the main analysis, individual-level data from the cohorts were analyzed in a single Cox model that was stratified by cohort. The model produced a single estimate for the disease of interest, but the effect estimates for adjusting variables (sex and the conventional potentially modifiable dementia risk factors) and the interactions between them were allowed to vary from cohort to cohort. The significant interactions were between sex and marital status, low education/socioeconomic status and physical inactivity, hypertension and marital status, obesity and smoking, and smoking and depression. We were only able to model the interaction between obesity and smoking for FPS and WHII, because HeSSup had no dementia cases among those who were both obese and smokers.
To examine the clustering of the index diseases that were associated with dementia, we computed odds ratios for the associations between these diseases by meta-analyzing data from the four cohorts and using the Mantel-Haenszel odds ratio method without continuity correction. 31 We used Stata MP 15 and 16 to analyze the data. The syntax for these analyses is available in the Appendix (Methods A.3). We report all confidence intervals at a 95% confidence level.   and duodenitis, and osteoporosis. The hazard ratios for these diseases ranged between 1.3 and 2.4 and were thus comparable to or higher than the relative risk estimates for conventional potentially modifiable dementia risk factors ( Figure 2 [ Table A.6 in the Appendix shows detailed data]). Unlike for the other five diseases, the hazard of dementia was not proportional for hypothyroidism, cerebral infarction, or osteoporosis, as the hazard ratio for dementia declined with increasing age ( Figure A.2 in the Appendix).

RESULTS
As shown in Figure 3, cerebral infarction and osteoporosis were more strongly associated with early and late-onset dementia (P for difference < .05), whereas the strength of the association for the remaining six physical diseases did not differ between early and lateonset dementia. For comparison, the associations between conventional risk factors and dementia did not differ between early and lateonset dementia (Table A.7 in the Appendix).

F I G U R E 3
Hazard ratio for early onset and late-onset dementia by exposure to eight physical diseases. ICD-10, International Classification of Diseases, 10th Revision. The bars represent 95% confidence intervals. Age is the time scale and hazard ratios are adjusted for sex. The within-study clustering of participants has been taken into account using cohort-specific baseline hazards and cohort-specific adjustment terms for sex.  The estimates from main analysis are the same as those from full dementia follow-up in Figure 1. a Additionally adjusted for the seven other physical diseases with robust long-term association with dementia. (There were two diseases reflecting coronary heart disease: myocardial infarction and chronic ischemic heart disease. To avoid over-adjustment, these two diseases were not adjusted for the other one.) effect on hazard ratios (Table A

DISCUSSION
In this exploratory multicohort study of over 280,000 adults, we quan- 5-year follow-up. 22 Earlier evidence of the association between hypothyroidism and dementia is also sparse. [32][33][34][35] A comparison of our findings on dementia-related cardiovascular disease to those of studies that have included clinical examinations to ascertain dementia supports the validity of our approach. In the full follow-up, our hazard ratios for acute myocardial infarction and chronic ischemic heart disease were 1.4 and 1.3, similar to the relative risk estimate of 1.3 for coronary heart disease and all-cause dementia in a recent meta-analysis. 16 For cerebral infarction, our hazard ratio was 2.4, whereas it was 2.2 in the 2018 systematic review and metaanalysis. 15 The present finding on osteoporosis is also consistent with previous studies reporting an association between low bone mineral density and dementia. 36,37 Several proposed pathological mechanisms may explain the associations between physical diseases and increased dementia risk. According to genetic studies, the immune system may play a role in the etiology of Alzheimer's disease. 20,21 Animal studies suggest that systemic inflammation can increase the blood-brain barrier's permeability, 38 induce inflammation into the central nervous system, 39 and cause long-term alterations in innate immune cells in the brain and an increased deposition of amyloid plaques, the hallmark of Alzheimer's pathology. 40 Sydenham's chorea is a specific example of a post-infectious autoimmune neurological disorder. It is a rare consequence of group A Streptococcus pharyngitis but is of interest here because erysipelas is also often caused by group A streptococci. 41,42 Inflammation may also contribute to increased dementia risk among those with duodenal ulcers, gastritis, and duodenitis. Helicobacter pylori infection is a shared cause of these diseases. 43,44 The evidence on the association between H. pylori infection and dementia is mixed 45 ; it is Plausible pathological mechanisms linking cardiovascular disease to dementia involve neuronal damage caused by microinfarcts, microbleeds, white matter lesions, and brain atrophy. 48 We found that myocardial infarction and chronic ischemic heart disease were associated with an increased risk of dementia, even after adjustment for hypertension, which is an established vascular risk factor for dementia. 1,3,4 In addition to direct neuronal damage, the association between cerebral infarction and dementia may reflect the impact of underlying systemic vascular disease. Little is known about the mechanisms underlying the association between hypothyroidism and dementia, although they could involve both direct effects on cognition and indirect effects through vascular pathology. 32,33,35 Our study has some important strengths. We investigated, in a single analytic setting, 22 physical diseases that have been associated previously with a higher incidence of dementia in the short-term (i.e., during the following 5 years), have been on a disease trajectory leading to dementia, or have been studied in separate studies with varying settings. 22 In contrast to earlier analyses, we also examined long-term associations that were less likely to be attributable to reverse causation and ascertainment biases.
Our study also has limitations and therefore should be considered hypothesis generating rather than conclusive. Although our main analysis of most diseases was well powered, the number of dementia cases among those exposed to the index diseases was low in some sensitivity analyses (e.g., occlusion and stenosis of precerebral arteries). For this reason, we may have missed diseases that are associated with an increased risk of dementia (Table A.9 in the Appendix). Given that the reproducibility of the findings was not confirmed in an independent study population, it also remains unclear whether our findings are replicable. We retrieved the information on the index diagnoses and incident dementia from electronic health records. This enabled all participants recruited for the study to be included in the analyses, rather than only those who continued to participate in follow-up examinations. However, although electronic health records have high positive predictive value (false positives are unlikely), they miss mild and undiagnosed cases of diseases. 49 This lack of sensitivity could contribute to the dilution of the results by nondifferential misclassification. The generalizability of our results is uncertain as the participants were mostly from occupational cohorts and therefore healthier than the general population. Our sample was relatively young, meaning a higher than usual proportion of early onset dementia cases. This may have led to an overestimation of the overall dementia risk related to cerebral infarction and osteoporosis, as these diseases had stronger associations with early than with late-onset dementia.
Due to reliance on observational data, our results may have been affected by confounding due to imprecise or lacking measurements of relevant covariates, such as conditions shared with the index disease (e.g., frailty) or off-target effects of medications used to treat these diseases.

CONCLUSION
In this exploratory study, the associations between eight physical diseases (erysipelas, hypothyroidism, myocardial infarction, ischemic heart disease, cerebral infarction, duodenal ulcer, gastritis and duodenitis, and osteoporosis) and dementia were equally strong or stronger than those between several currently targeted risk factors (low education, hypertension, obesity, smoking, depression, physical inactivity, social isolation, and diabetes) and dementia. Further research on the mechanisms underlying these disease-dementia associations is needed to examine whether the observed associations are causal and to confirm the relevance of our findings for dementia etiology and prevention.

ACKNOWLEDGMENTS
The study was supported by grants from NordForsk, the Academy of the collection, management, analysis, or interpretation of the data; the preparation, review or approval of the manuscript; or the decision to submit the manuscript for publication.