Association of hospitalization with long-term cognitive and brain MRI changes in the ARIC cohort

Abstract

Objective:

To determine whether hospitalization is associated with subsequent cognitive decline or changes on brain MRI in a community-based cohort.

Methods:

Baseline and follow-up cognitive testing (n = 2,386) and MRI scans with standardized assessments (n = 885) were available from a subset of white and black participants in the Atherosclerosis Risk in Communities study. Cognitive tests included the Delayed Word Recall Test (DWRT), Digit Symbol Substitution Test (DSST), and Word Fluency Test (WFT). Hospitalization characteristics were determined using ICD-9 codes. Regression models adjusted for demographics, education, comorbidities, and APOE ε4 were used to estimate the independent association of hospitalization with changes in cognition or neuroimaging.

Results:

Over a mean 14.1 years between visits, 1,266 participants (53.1%) were hospitalized. Hospitalization compared with no hospitalization was associated with greater decline in DSST scores (1.25 points greater decline, p < 0.001) but no difference in DWRT or WFT score change. Each additional hospitalization, as well as a critical illness vs noncritical illness hospitalization, was associated with greater decline in DSST scores. A subset of participants (n = 885) underwent MRI scans separated by 10.5 years. Hospitalization (n = 392) compared with no hospitalization was associated with a 57% higher odds of increasing ventricular size at follow-up. Each additional hospitalization, as well as having a critical illness vs noncritical illness hospitalization, and having a hospitalization with major surgery vs no surgery was associated with greater odds of increased ventricular size.

Conclusions:

Cognitive decline and neuroimaging changes may occur after hospitalization, independent of baseline demographics and comorbidities.

Hospitalizations in older adults (>65 years) account for more than one third of community hospital stays and result in substantial morbidity and mortality.^1–3 Recently, evidence has emerged that hospitalization may be associated with long-term cognitive decline and incident dementia, even in the absence of hospitalization for stroke or head trauma.^4–6 However, the association between hospitalization and cognitive decline has not been well-characterized, although establishing the magnitude and etiology of cognitive decline would be crucial to develop targeted prevention strategies.

In order to characterize the independent associations of hospitalization with cognitive change, accurate assessment of a patient's prehospitalization comorbidities is critical, since hospitalization might only be a marker of vulnerability. However, previous studies have been limited by excessive reliance on patient self-reported comorbidities, which may lead to misclassification with residual confounding, and lack of accounting for APOE ε4 status. The potential mechanism for cognitive decline following hospitalization for reasons other than stroke is unknown, although both patient and hospitalization characteristics may provide insight. Importantly, changes on brain imaging, including incident infarcts,⁷ brain atrophy,⁸ and increases in white matter hyperintensities (WMH)^9,10 have all been associated with cognitive decline and may be more common after hospitalization.

The Atherosclerosis Risk in Communities (ARIC) study is a predominantly biethnic longitudinal cohort study with baseline and follow-up cognitive testing 15 years apart, baseline and follow-up brain MRI 10 years apart, and rigorous characterization of patient comorbidities. In this study, the association between hospitalization and cognitive decline and neuroimaging changes is examined in the ARIC cohort.

METHODS

Study population.

The ARIC study is a prospective, population-based cohort study conducted in a predominantly biethnic cohort of 15,792 participants enrolled between 1987 and 1989. Study design and procedures have been published.¹¹ All participants were invited for assessments at visits 2, 3, and 4, while only a subset of participants were invited for brain or carotid MRI visits. Briefly, the brain MRI cohort was >55 years old from Forsyth County, North Carolina, or Jackson, Mississippi, without exclusions for MRI, invited as an ancillary to visit 3 (1993–1995).^12,13 The carotid MRI cohort was invited based on a stratified sampling plan from all sites, with a goal of recruiting 1,200 participants with high values of carotid arterial intimal medial thickness (IMT) and 800 individuals randomly sampled from the remainder of the IMT distribution.¹⁴

For the purpose of examining cognitive decline, we limited the analyzed population to participants who had cognitive testing at visit 2 (1990–1992, baseline) and either the brain¹² or carotid MRI visits¹⁴ (2004–2006, follow-up), since these visits provided the largest interval of cognitive testing.

For the purpose of examining changes on brain MRI, we limited the analyzed population to participants who had brain MRI scans at visit 3 (1993–1995, baseline) and the brain MRI visit (2004–2006, follow-up).

For all analyses, we excluded participants with a primary brain injury (ischemic or hemorrhagic stroke or head trauma) noted during hospitalization between baseline and follow-up, and patients on antidementia, anticonvulsant, or antipsychotic medications at baseline. We did include patients hospitalized with conditions (besides stroke or head trauma) affecting the CNS. Participants missing baseline or follow-up cognitive testing or covariate data were excluded, and we disregarded hospitalizations in the year before follow-up visits (to reduce short-term effects of hospitalization). Figure 1 shows included patients.

Figure 1. Flow diagram.

Patients included and excluded for analysis of cognitive and brain MRI changes in the Atherosclerosis Risk in Communities study.

Standard protocol approvals, registrations, and patient consents.

All participants provided written informed consent, and the study was conducted with institutional review board approval at all study sites.

Measures of hospitalization.

The ARIC study obtains hospitalization information through annual telephone contact with participants and active surveillance of local hospitals. These data include all ICD-9 codes. We categorized each hospitalization as noncritical or critical illness, using the methodology of Ehlenbach et al.⁴ with the addition of sepsis (ICD-9 code 995.91) as a critical illness. Critical illness included shock/hypotension, sepsis, respiratory failure or prolonged mechanical ventilation, and respiratory/cardiac arrest. We further categorized each hospitalization into major surgical or nonsurgical using Clinical Classification Software (CCS) (http://www.hcupus.ahrq.gov/toolssoftware/ccs/ccsfactsheet.jsp) from the Agency for Healthcare Quality and Research. We defined major surgical procedures as those with intermediate or high cardiac risk according to guidelines,¹⁵ generally spine, thyroid, cardiothoracic, vascular, gastrointestinal, hernia, or genitourinary procedures. We defined nonsurgical hospitalizations as not having CCS procedure codes (except radiologic imaging).

Outcome measures of cognitive function.

Three cognitive tests were administered: (1) Delayed Word Recall Test (DWRT)¹⁶ (test of verbal learning and recent memory); (2) Digit Symbol Substitution Test (DSST)¹⁷ (test of executive function and processing speed; and (3) Word Fluency Test (WFT)¹⁸ (test of executive function and expressive language). For participants with cognitive testing at both the brain and carotid MRI visits, only the first score was used due to potential practice effects.¹⁹ The primary cognitive outcome was change in each cognitive test score, obtained by calculating visit 2 raw score minus the MRI visit raw score.

MRI measurement and outcomes.

Details of the visit 3 and brain MRI visit scanning/interpretation protocols have been published.^13,20,21 Infarcts were defined based on signal characteristics: bright on T2 and proton density scans and dark on T1 images. Ventricular size and sulcal width were estimated from axial T1-weighted images using a validated 0–9 scale.²¹ Outcomes of interest were presence of any new infarct, worsening of ≥1 grade of ventricular size (0–9 scale), and worsening of ≥1 grade of sulcal size (0–9 scale).

Change in WMH volume from visit 3 to the brain MRI visit was calculated by subtracting the WMH volume at the brain MRI visit²⁰ (using quantitative WMH volumes) minus the estimated WMH volume at visit 3 (using a validated prediction equation that related WMH volume to visit 3 visual grade) (0–9 scale) (R² = 0.8).²⁰ For WMH volume change, which was categorized into quintiles because of non-normal distribution, the outcome was odds of being in the top quintile of change compared to the bottom 4 quintiles, using prior methodology.²⁰ We also examined odds of being in the top quintile of WMH visual grade increase.

Statistical analysis.

Characteristics of patients were compared using t tests or Wilcoxon rank-sum tests for continuous variables, χ² tests for binary variables, and Wilcoxon rank-sum tests for categorical variables.

Cognitive change analysis.

First, multivariable linear regression models were used to estimate the change in each cognitive test score, comparing hospitalized vs nonhospitalized participants in the visit 2–MRI visit interval. Second, among participants with first hospitalization after visit 2, we used multilevel linear regression models with patient-level random intercepts, considering raw cognitive scores at visits 2, 3, 4, and MRI visit. For each subject, we defined time 0 as time of first hospitalization and calculated time from first hospitalization based on the dates of each visit. We estimated the change in raw cognitive scores before and after first hospitalization by including the main effect of time to estimate the slope prior to first hospitalization and a linear spline term that allowed the slope for time to change after the first hospitalization.

Neuroimaging analysis.

MRI outcomes were dichotomized (increase of ≥1 grade of ventricular, or sulcal size, any new infarct, being in the top quintile of WMH volume change, or increase in visual grade). Multivariable logistic regression models were used to estimate the odds of each of these outcomes, comparing hospitalized vs nonhospitalized participants in the visit 3–MRI visit interval.

Secondary analyses.

Additional exploratory characterizations of hospitalization were separately evaluated, including (1) number of hospitalizations (0, 1, 2, 3, 4, 5, ≥6), (2) critical illness vs noncritical illness hospitalization, (3) major surgical vs nonsurgical hospitalization, and (4) black vs white race. A potential role for neuroimaging changes on mediating cognitive changes was examined by adding neuroimaging changes to the adjusted model.

All models included age, sex, race/center, education (<high school, high school degree or vocational school, >high school), hypertension (diastolic blood pressure >90 mm Hg or hypertension medications), systolic blood pressure, coronary heart disease (self-reported at visit 1 or subsequent adjudicated events), smoking status (never vs ever), diabetes (fasting plasma glucose >126 mg/dL or any nonfasting glucose >200 mg/dL, use of diabetes medications, or self-reported), self-reported or adjudicated stroke, carotid intimal medial thickness, number of APOE ε4 alleles (defined using single nucleotide polymorphisms rs429358 and rs7412 and coded 0, 1, or 2 alleles), low-density lipoprotein (LDL) concentration, kidney disease (estimated glomerular filtration rate <60 mL/min/1.73 m² using the CKD-EPI equation²²), and the following self-reported diseases at visit 4: lung disease, hepatitis, lupus, rheumatoid arthritis, and cancer. To account for missing covariate data as a sensitivity analysis, we used multiple imputation by chained equations to create 5 imputed datasets using the missing data module in Stata.²³ Parameter estimates were averaged across imputed datasets. Standard errors of the estimates were calculated using Rubin's formula.²⁴

All p values are 2-sided, and p < 0.05 was considered statistically significant. Stata version 12 (StataCorp, College Station, TX) was used for analyses.

RESULTS

Participant characteristics.

For analysis of cognitive change, data were analyzed for 2,386 participants. Over a mean of 14.1 ± 1.0 years between baseline and follow-up visits, 1,266 participants (53.1%) were hospitalized at least once. For analysis of neuroimaging changes, data were analyzed for 885 participants. Over a mean of 10.5 ± 0.8 years between baseline and follow-up neuroimaging, 392 patients (44.3%) were hospitalized at least once. Figure 1 shows a flowchart of included patients.

Baseline characteristics of the study populations, as well as the overall ARIC population, are shown in table 1. Hospitalized participants were generally older, were more likely to be men, and were more likely to have more carotid atherosclerosis and comorbidities, including hypertension, diabetes, coronary heart disease, kidney disease, and lung disease. Baseline cognitive scores were slightly lower among hospitalized participants. Brain MRI characteristics at baseline were generally similar, although there were more prevalent infarcts among hospitalized participants.

Table 1.

Participant characteristics

Hospitalization and cognitive change.

In adjusted models (table 2), there was more decline in DSST scores among participants with a hospitalization between ARIC visit 2 and the MRI visits, compared to participants without hospitalization in that interval (greater decline by 1.25 symbol pairs) However, the adjusted decline in both WFT and DWRT scores during the visit 2–MRI visits interval was not different among hospitalized vs nonhospitalized participants.

Table 2.

Relative decline in cognitive test scores for hospitalized vs nonhospitalized participants and other exploratory characterization of hospitalizations

Because of baseline differences in participants, we also examined trajectories of cognitive change before hospitalization compared to after hospitalization. Before hospitalization, the average rate of DSST decline was 0.38 digit-symbol pairs/year (95% confidence interval [CI] 0.34–0.42), which after hospitalization worsened by an additional 0.20 digit-symbol pairs/year (95% CI 0.12–0.27; p < 0.001). Similarly, the average rate of WFT decline before hospitalization was 0.10 words/year (95% CI 0.05–0.14), which after hospitalization worsened by an additional 0.09 words/year (95% CI 0.02–0.17, p = 0.02). There was no statistical difference in average rate of DWR decline before hospitalization (0.032 words/year [95% CI 0.023–0.041]) compared to after hospitalization (worsening by 0.016 more words/year [95% CI −0.001 to 0.032; p = 0.06]). Inferences regarding cognitive change were not different when multiple imputation was used to account for missing covariate data.

Hospitalization and changes on MRI studies.

As shown in table 3, the odds of ventricular size increasing from baseline to follow-up were 1.57-fold greater among hospitalized compared to nonhospitalized participants. The odds of sulcal size increasing were also greater, although not significant (p = 0.06), among hospitalized participants, and there was no difference in incident infarcts by hospitalization status.

Table 3.

Adjusted odds of neuroimaging outcomes from visit 3 to the MRI visits by hospitalization status, and other exploratory characterization of hospitalizations

The odds of being in the top quintile of WMH volume change compared to the bottom 4 quintiles were 1.55 times higher for hospitalized compared to nonhospitalized participants. However, when we examined change in WMH using visual grade¹⁶ (0–9 scale) to account for potential imprecision in continuous volumetric estimates, we found no difference in the odds of being in the top quintile of increase in WMH visual grade by hospitalization status. Inferences were unchanged when multiple imputation was used to account for missing covariate data.

Exploratory characterizations of the hospital exposure.

Number of hospitalizations.

There was a general dose effect in the association of each additional hospitalization and decline in DSST scores (0.58 symbol-digit pair decline with each hospitalization) and WFT scores (0.27 symbol-digit pair decline with each hospitalization) (figure 2 and table 2). With regard to neuroimaging outcomes, each additional hospitalization was associated with increased odds of being in the top quintile of WMH volume progression (odds ratio [OR] 1.23), of incident infarct (OR 1.17), and of ventricular size increase (OR 1.21) (table 3).

Figure 2. Adjusted decline in cognitive test scores by number of noncritical illness hospitalizations.

Adjusted decline in test scores for participants by number of hospitalizations, compared to the reference group of test score decline in participants with no hospitalizations. A (+) value for decline indicates that test scores worsened between baseline and follow-up, while a (−) value for decline indicates that test scores improved between baseline and follow-up. *Above each bar indicates a p value <0.05 for the comparison of cognitive decline for that particular number of hospitalizations compared to the reference of no hospitalizations. Overall p value refers to the overall association between each additional hospitalization and decline in individual test score.

Severity of hospitalization.

The number of participants with critical illness was small (n = 28 for cognitive change and n = 10 for neuroimaging), thus limiting the ability to make rigorous comparisons. Nevertheless, there was significantly greater decline in DSST scores for participants hospitalized with critical compared to noncritical illness (table 2). With regard to neuroimaging outcomes, ventricular size increased by at least 1 grade in every participant hospitalized with critical illness, and this increase was significant compared to ventricular changes in participants with noncritical illness hospitalization (p = 0.02) (table 3).

Major surgical vs nonsurgical hospitalization.

There were no large differences in cognitive outcomes among participants hospitalized with major surgery compared to no surgery (table 2). With regard to neuroimaging outcomes (table 3), the odds of worsening ventricle size were significantly increased among participants hospitalized with major surgery compared to no surgery (table 3).

Effect of race.

Changes in cognitive and neuroimaging outcomes were examined according to race, and there was no significant race interaction for the association of hospitalization with any outcome (all p interactions >0.05).

Potential mediation of cognitive changes by neuroimaging outcomes.

We examined whether changes in ventricular size, white matter, or infarcts from the visit 3–MRI visit mediated the decline in DSST scores that was observed after hospitalization from visit 2–MRI visit. We found a 25% attenuation of estimated change in DSST scores when the amount of ventricle change was added to the model, indicating a modest mediating effect on decline in DSST scores. We found only small attenuation of cognitive coefficients (<15%) with the addition of WMH volume change, WMH visual grade increase, or incident infarcts to the model, indicating no substantial mediating effect.

DISCUSSION

In this study of a subset of ARIC participants, hospitalizations were independently associated with a decline on one test of executive function (DSST), similar to results of previous studies,^4,5 even with inclusion of more precise comorbidity measures. Further, hospitalizations were associated with increased ventricular size on brain MRI.

There are many potential explanations for the observed cognitive decline. Events during hospitalization have been associated with cognitive decline, including delirium,^25–27 hypotension,²⁸ sedation,²⁹ acute lung injury,³⁰ and sepsis.³¹ Older adults may be particularly susceptible because of less cognitive reserve or increased susceptibility to inflammation.³² To better understand the etiology of decline in DSST scores, we examined characteristics of hospitalizations and found that both number of hospitalizations and critical illness were associated with increased decline in DSST scores, implying a potential dose effect of insults during hospitalization. There was no difference in DSST score change by race, or by nonsurgical vs major surgical hospitalization, implying that factors driving cognitive change may not be unique to the perioperative period.

We were particularly interested in the effect of hospitalization on changes in brain MRI, since we hypothesized that hospitalization might lead to structural brain changes and subsequent cognitive decline. Our results demonstrate increased ventricular size after hospitalization compared to no hospitalization, with both critical illness and major surgical hospitalizations even more highly associated with worsening ventricular size. These results are consistent with prior studies that have demonstrated associations between events during hospitalization (acute respiratory distress syndrome³³ or surgery³⁴) and increased cerebral atrophy, but the length of follow-up and the adjustment for important covariates distinguish our results.

We examined the potential mediating role of changes in ventricular size or WMH on cognitive decline, since both cerebral atrophy^35,36 and WMH severity³⁷ and progression^38,39 have been associated with decreases in cognition. Our results suggest that increased ventricular size (a marker of cerebral atrophy), but not changes in WMH, may mediate a portion of decline in DSST scores, although it appears that a large portion of cognitive change occurs independent of structural changes observable on brain MRI. However, these results are limited because MRI changes were assessed from visit 3–MRI visits, while cognitive change was assessed from visit 2–MRI visits, and thus should be considered hypothesis-generating.

An alternative explanation for our results is that underlying comorbidities may predispose patients to both hospitalizations and further cognitive or MRI changes. To account for this possibility, previous studies have adjusted for patient self-reported comorbidities, thus leading to potential misclassification bias. We adjusted our models using precise comorbidity definitions or actual physiologic or genetic data from a well-characterized cohort. Additionally, we examined cognitive trajectories before and after hospitalization with each participant serving as his or her own control. Similar to our main results, we found worsening trajectory of decline in cognitive scores after hospitalization compared to before hospitalization.

Strengths of this study include a well-defined cohort with precise comorbidity definitions and both cognitive and MRI testing separated by ≥10 years. However, there are several limitations. First, as an observational study, our results could be explained by residual unmeasured confounding. Second, analysis of potentially important hospitalization characteristics may be underpowered to detect a difference among groups, and other factors (besides critical illness or major surgery) might be important in cognitive or brain MRI changes. Additionally, the association of white matter change with hospitalization was only significant when white matter change was considered as a continuous (not categorical) variable. Third, because the study population was slightly healthier than the general ARIC population, results may not be generalizable to all patients. Finally, cognitive testing was limited to 3 tests, which may not adequately characterize cognitive change. Recently, ARIC participants underwent extensive neurocognitive testing and dementia classification, so further studies in ARIC participants will be able to address this limitation.

The results of this study highlight cognitive consequences of hospitalization in a community cohort. The results provide a rationale for research to characterize the timing and etiology of cognitive and brain MRI changes after hospitalization, to determine modifiable risk factors, and to determine whether increased prevention, monitoring, or rehabilitation efforts can improve cognitive outcomes after hospitalization.

GLOSSARY

ARIC

Atherosclerosis Risk in Communities

CCS

Clinical Classification Software

CI

confidence interval

DSST

Digit Symbol Substitution Test

DWRT

Delayed Word Recall Test

ICD-9

International Classification of Diseases–9

IMT

intimal medial thickness

LDL

low-density lipoprotein

OR

odds ratio

WFT

Word Fluency Test

WMH

white matter hyperintensities

AUTHOR CONTRIBUTIONS

Charles Brown helped design the study, perform the statistical analysis, draft and review the manuscript, and approved the final manuscript. A. Richey Sharrett helped design the study, revise the manuscript, and approved the final manuscript. Josef Coresh provided input into the analysis, helped revise the manuscript, and approved the final manuscript. Andrea L.C. Schneider provided input into the statistical analysis, helped revise the manuscript, and approved the final manuscript. Alvaro Alonso provided input into the analysis, helped revise the manuscript, and approved the final manuscript. David S. Knopman provided input into the analysis, helped revise the manuscript, and approved the final manuscript. Thomas H. Mosley provided input into the analysis, helped revise the manuscript, and approved the final manuscript. Rebecca F. Gottesman helped design the study, provided input into the statistical analysis, helped revise the manuscript, and approved the final manuscript.

STUDY FUNDING

Andrea Schneider was supported by NHLBI grant 5T32HL007024. Charles Brown was supported by R01 AG033615-01A1 NIH KL-2 Clinical Research Scholars Program, RO3 AG042331, and Jahnigen Career Development Award. ARIC is conducted as a collaborative study supported by National Heart, Lung, and Blood Institute Contracts HHSN268201100005C, HHSN268201100006C, HHSN268201100007C, HHSN268201100008C, HHSN268201100009C, HHSN268201100010C, HHSN268201100011C, HHSN268201100012C, R01HL087641, R01HL59367, and R01HL086694; National Human Genome Research Institute contract U01HG004402; and National Institutes of Health contract HHSN268200625226C. ARIC carotid MRI examination is funded by U01HL075572-01 and the ARIC brain MRI examination is funded by grant R01-HL70825. Neurocognitive data are collected by U01 HL096812, HL096814, HL096899, HL096902, and HL096917. Infrastructure is partly supported by grant UL1RR025005, a component of the NIH and NIH Roadmap for Medical Research. This study was not industry-sponsored.

DISCLOSURE

C. Brown, A. Sharrett, J. Coresh, A. Schneider, and A. Alonso report no disclosures relevant to the manuscript. D. Knopman serves as Deputy Editor for Neurology® and serves on a Data Safety Monitoring Board for Lundbeck Pharmaceuticals and for the Dominantly Inherited Alzheimer's Disease Treatment Unit; served on a Data Safety Monitoring Board for Lilly Pharmaceuticals; served as a consultant to Tau RX; was an investigator in clinical trials sponsored by Baxter and Elan Pharmaceuticals in the past 2 years; and receives research support from the NIH. T. Mosley and R. Gottesman report no disclosures relevant to the manuscript. Go to Neurology.org for full disclosures.