How Does Dementia Begin to Manifest in Bipolar Disorder? A Description of Prodromal Clinical and Cognitive Changes

BACKGROUND
Older adults with bipolar disorder (BD) have increased dementia risk, but signs of dementia are difficult to detect in the context of pre-existing deficits inherent to BD.


OBJECTIVE
To identify the emergence of indicators of early dementia in BD.


METHODS
One hundred and fifty-nine non-demented adults with BD from the National Alzheimer Coordinating Center (NACC) data repository underwent annual neuropsychological assessment up to 14 years (54.0 months average follow-up). Cognitive performance was examined longitudinally with linear mixed-effects models, and yearly differences between incident dementia and controls cases were examined in the six years prior to diagnosis.


RESULTS
Forty participants (25.2%) developed dementia over the follow-up period ('incident cases'). Alzheimer's disease was the most common presumed etiology, though this was likely a result of sampling biases within NACC. Incident cases showed declining trajectories in memory, language, and speeded attention two years prior to dementia onset.


CONCLUSION
In a sample of BD patients enriched for Alzheimer's type dementia, prodromal dementia in BD can be detected up to two years before onset using the same cognitive tests used in psychiatrically-healthy older adults (i.e., measures of verbal recall and fluency). Cognition in the natural course of BD is generally stable, and impairment or marked decline on measures of verbal episodic memory or semantic retrieval may indicate an early neurodegenerative process.

Despite this apparent heightened risk, studies aiming to increase knowledge about dementia syndromes systematically exclude subjects with psychiatric conditions (12). In addition, the majority of published studies examining cognition in adult BD patients include only subjects younger than age 50 (13). Thus, relatively little is known about cognitive decline in BD later in life, or how clinicians should go about evaluating cognitive performance in the context of pre-existing de cits. Confusion between the 'core' de cits (i.e., de cits that are part of the natural history of BD) and those that signal a neurodegenerative process has been highlighted previously (14,15). Therefore, establishing features signaling the onset of neurodegenerative changes in BD would be helpful to inform the differential diagnosis at a critical period when early therapeutic intervention may be possible and important healthcare decisions must be made.
Previous studies have called for the identi cation of neuropsychological markers of prodromal dementia in individuals with BD (1,2). The few existing investigations of cognitive symptoms in older cohorts have focused on comparing elderly individuals with BD to age-matched controls without BD who are cognitively healthy (e.g., 16,17) or cognitively impaired (e.g., 18). Only one previous study has directly compared cognitive performance in BD patients with and without dementia (2); in this study, the authors assessed the use of the Cambridge Cognition Examination, Mini-Mental State Examination (MMSE), animal uency, and clock drawing to distinguish BD with mild AD from BD alone, and reported good sensitivity and speci city for all tests. However, it remains unclear whether or not cognitive measures have value to identify BD subjects who are in the earliest stages of a neurocognitive disorder given that all participants in the aforementioned study already had dementia. Furthermore, participants with non-AD dementias were excluded.
Here, we used longitudinal neuropsychological data over a 10-year span to document the emergence of early features of dementia in BD, relative to those that constitute the natural course of cognitive changes in individuals with BD who do not develop dementia. The approach was exploratory and data-driven, and therefore did not include any a priori hypotheses.

Participants
This retrospective observational study used data from the National Alzheimer's Coordinating Center (NACC; www.alz.washington.edu). The database consists of participants recruited from several Alzheimer's Disease Centers (ADC) across the USA, and includes individuals with dementia, mild cognitive impairment (MCI) or normal cognition. Data for this study were for visits conducted between August 2005 and the September 2016 data freeze, and included any participant with self-reported or cliniciandiagnosed BD.
Participants were 241 adults with BD seen at approximately yearly intervals. Individuals with no follow-up after their rst visit (n=81) or those who were demented on admission (n=75) were excluded; 29 participants met both exclusion criteria. The remaining 114 participants were seen up to 112 months (4.3 total visits including baseline, or 42.8 months, on average). Sixty-four percent had self-reported BD, with the remaining 36% being clinician-diagnosed. Participants were classi ed as 'incident dementia cases' if they developed dementia at any point during the study. Dementia diagnoses were based on clinical judgement using DSM-IV criteria (19) (or similar/modi ed criteria; NACC acknowledges that diagnostic criteria may have varied between ADC sites). DSM-IV criteria include: 1) de cits in memory and at least one other cognitive domain; 2) signi cant impairment in social or occupational functioning due to cognitive impairments; 3) gradual onset and progressive course; 4) absence of contributing neurological or systemic conditions, delirium or other mental disorder. Twenty-six participants (22.8%) developed dementia during the follow-up period; the remaining 88 cases (77.2%) remained free of dementia for the duration of the study (up to nine years) and constituted the control group.
All contributing ADCs obtained participant consent and received approval from their individual Institutional Review Board prior to submitting data to NACC.

Risk factors for cognitive decline
The presence of neuropsychiatric symptoms (20) and vascular risk factors (21) has been associated with cognitive decline in non-BD populations. To determine the impact of these factors on dementia risk in this sample, data were used from the Neuropsychiatric Inventory (NPI) (22), which was completed by a knowledgeable informant. Vascular risk factors were assessed using the Hachinski scale (23). Information about tobacco use (total years smoked) and alcohol abuse (absent, recent/active, remote/inactive) was collected. Mood-stabilizing medications, such as lithium, are also known to impact cognition (24). Medication use within two weeks of assessment was documented in the present sample; unfortunately, lifetime exposure was not. All mood-stabilizing medications, as well as antipsychotics, benzodiazepines, and other antiepileptic drugs that can be used off-label as mood stabilizers, were considered.

Cognitive outcomes
All participants were administered the ADC Uniform Data Set (UDS) neuropsychological battery (25), which included the Wechsler Memory Scale Revised (WMS-R) forward and backward digit spans, WMS-R digit-symbol substitution (DSS), trails A and B, WMS-R Logical Memory Story A, semantic uency (animals and vegetables) and confrontation naming. The MMSE was used as a brief measure of general cognition. Variables coded as 'missing due to a cognitive/behavior problem' were replaced with the lowest allowable score for that test. All other missing variables were left missing. Scores were then standardized to Z-scores using normative data derived from 3,268 cognitively normal participants from NACC published previously (26), where Z=0 provides a benchmark for expected performance in healthy older adults (with an associated standard deviation of 1), thus precluding the need for a non-psychiatric normal control group. In addition to formal cognitive testing, participants were asked about subjective reports in memory decline relative to previous abilities (one subject had missing data for this question). They were also queried about the predominant cognitive symptom that was rst recognized as a change from previous levels.

Structural neuroimaging
The NACC data repository included available structural magnetic resonance imaging (MRI) data on a small subset of 12 BD participants, including 5 incident dementia cases (4 males, 1 female) and 7 controls (4 males, 3 females). These images were analyzed to obtain supplementary insights into the possible neurological contributions to cognitive decline in BD. To this end, intracranial volumes were acquired using automatic segmentation in SPM12 (www. l.ion.ucl.ac.uk/spm) which calculated white matter, grey matter, and cerebral spinal uid volumes for each participant. These volumes were summed and corrected for voxel size to ensure consistency across participants. In addition, because AD is the most frequent type of dementia seen in BD (1) and is characterized by signi cant hippocampal atrophy, bilateral, manual segmentations of the hippocampi were traced for each participant using ITK-SNAP (27) (www.itksnap.org). A well-established protocol for manual segmentation of the hippocampus in T1weighted images was used to train both the primary and secondary rater (28). Tracing reliability was assessed by having a second rater trace four of the 12 brains (two from each group). Volumes from the four tracings were highly correlated between raters (r (2) = .926, p = .247) suggesting a high degree of inter-rater reliability. Left and right hippocampal volumes were summed and corrected for voxel size to produce total hippocampal formation volumes for each participant.

Statistical analyses
All analyses were two-tailed, with alpha set at .05 except where stated otherwise. The incident dementia cases and dementia-free controls were rst compared in terms of age and years of education using Mann-Whitney U, because the data were non-normally distributed (Shapiro-Wilk p<.001 for both variables). Sex was compared between groups using chi-square (χ 2 ).
Neuropsychiatric symptoms, vascular risk factors, tobacco, alcohol and medication use, and subjective impression of cognitive change at any time during the study (yes/no) were compared longitudinally using linear mixed models with Group (incident cases, controls) as a xed effect. Because all medications (n=29) were considered independently in separate statistical analyses, the signi cance threshold was set at p < .002 using Bonferroni correction for multiple comparisons. A Group × Time interaction term was added to the model with neuropsychiatric symptoms in order to determine whether worsening neuropsychiatric symptoms were associated with incident dementia.
Cognitive change was examined as a function of the time interval (in months) between each assessment and 'Time0', ranging from -112 to 0. 'Time0' was always the date of dementia diagnosis for incident dementia cases, and last follow-up visit for controls. Of note, the earliest date of dementia diagnosis was always used, even if diagnosis subsequently reverted to MCI (n=5; two of these ultimately were rediagnosed with dementia at a later visit). Linear mixed models were used to establish the relationship between longitudinal performance on each cognitive test and dementia status. This statistical approach allows for correlated repeated measures within subjects, while permitting the inclusion of participants with missing data at certain time points (29). All longitudinal analyses included only data collected between Time-6 and Time0 because only two incident dementia cases had data prior to Time-6. Cognitive variables were entered individually into separate models with Time and Group as xed effects, and covarying for sex as a xed effect use due to a priori group differences (see Results, below). Group × Time interaction terms were added to all models and were the main outcome of interest. Six participants were excluded from these analyses because they lacked cognitive data at all time points, which were coded as 'Missing due to: Other problem' (i.e., not due to a physical, cognitive or behavioral problem, nor due to verbal refusal).
Hippocampal atrophy was examined using a one-way analysis of covariance (ANCOVA) accounting for age, intracranial volume, and the number of days between Time0 and the date of MRI acquisition. Additionally, a group effect was discovered in the number of days between Time0 and the date of MRI acquisition. Consequently, group means were removed from this variable by subtracting the group mean from the amount of days to Time0 for each participant. This was done to remove any inter-group variance while still maintaining a marginal amount of intra-group variance. Sex was not included as a covariate in these analyses because there were no sex differences in the neuroimaging sample (χ 2 =0.686, p=0.408).

Participant characteristics
Participants' sociodemographic characteristics are summarized in Table 1. The 26 incident cases developed dementia after a mean of 33.96 months (SD=25.71), and the 88 controls remained free of dementia for a mean follow-up duration of 45.39 months (SD=27.46). Follow-up length was signi cantly longer in controls (χ 2 =4.955, p=.026). Age at Time0 (χ 2 =2.970, p=.085) and education (χ 2 =3.325, p=.068) were comparable between groups. The incident dementia group included a signi cantly greater proportion of males (χ 2 =14.954, p<.001). A signi cantly greater proportion of incident dementia cases reported concern about their memory compared to controls (χ 2 =4.769, p=.029). BD was determined to be the primary etiology contributing to cognitive impairment in two participants (7.7%), one of whom also had depression and Parkinson's disease as primary contributing factors. Fifteen participants (57.7%) had AD as a primary etiologic diagnosis, three (11.5%) had fronto-temporal lobar degeneration, two (7.7%) had Lewy body dementia. One participant had alcohol-related dementia, one had vascular dementia and one had normal-pressure hydrocephalus. In one participant, etiology was undetermined. These etiologic diagnoses were presumptive and determined by clinicians within individual ADCs based on clinical presentation.
Sex-adjusted linear mixed model results are illustrated in Figure 1. To determine the earliest point at which between-group differences appeared in models yielding a signi cant interaction term, yearly differences between incident cases and controls were examined using sex-adjusted ANCOVA. Only data from Time-6 to Time0 were included, due to insu cient data at prior time points. Because DSS and Logical Memory were the rst to be affected in the incident dementia group, their usefulness as potential screening tools for dementia in BD was tested in exploratory post-hoc analyses. The presence of impairment (yes/no) on the DSS or Logical Memory immediate story recall at any time during the study (prior to their last visit) was rst determined for each participant. Impairment was de ned as performance below -1.5 SD based on published normative data (26). Sensitivity, speci city,

Discussion
To distinguish cognitive de cits inherent to BD from those signaling incident dementia, this study examined a large cohort of older adults with BD. Over the course of the study, nearly a quarter (22.8%) of the eligible sample eventually received a diagnosis of dementia, within the range of previously-reported dementia rates of 20-25% in BD (9,10) and higher than the 7% population prevalence in non-psychiatric samples (11). This group included predominantly males (76.9%), and most endorsed subjective concerns at some point during the study. Memory problems were the predominant symptoms rst recognized as a change in cognition, and AD was the most commonly diagnosed etiology, consistent with prior reports in BD (1) and supported by between-group differences in hippocampal volumes in a small subset of participants with available structural neuroimaging data.
In this study, we showed for the rst time that a standard neuropsychological assessment can be used to detect signs of dementia up to ve years prior to dementia diagnosis in individuals with BD, and that neurodegenerative cognitive processes in this group of patients resemble that of the general population. Though this nding may appear self-evident, to our knowledge, it has not been documented empirically by previous studies. We found that performance on all cognitive measures was associated with incident dementia, with short story immediate recall and DSS showing the earliest between-group differences, up to four and ve years before diagnosis, respectively. Importantly, DSS and story recall were not impaired in BD controls who remained dementia-free throughout the study: at all time points, these patients' performance hovered around Z=0 (i.e., clinically normal) for both measures (Figure 1). Although this contrasts with reports of cognitive impairments being a core feature of BD (1), it is in line with other evidence suggesting that cognitive impairments within the natural course of BD are mostly mild (3,5,6).
Story recall has been identi ed in other studies as being sensitive to medial temporal lobe atrophy, the central neuroanatomical feature of AD (32). Regarding DSS, a measure of psychomotor speed that also taps sustained attention and set-shifting, other studies have similarly reported that it may be a particularly sensitive early measure of incident dementia (33,34), up to ten years before diagnosis (35), potentially re ecting its association with medial temporal lobe atrophy severity and cerebrovascular burden (36). In this sample, the value of these measures lay primarily in their negative predictive power, i.e., to rule out a possible dementing condition if performance is normal. If independent studies replicate our ndings in broader BD samples, these measures may be worthwhile considering as potential costeffective clinical screening tools for dementia risk in BD in memory clinics. It may also be worthwhile systematically investigating the use of these screening tools in other patient populations affected by memory disorders (e.g., schizophrenia: ,37).
Factors such as diminished cognitive reserve, vascular risk factors, and exposure to pharmacological treatments have been proposed to account for increased dementia risk in BD (7). Although the present study was not designed to speci cally test these associations, its results do not appear to support this view. No differences were found between the incident cases and controls cases in terms of education (considered a cognitive reserve proxy: ,38), or Hachinski score (a measure of vascular risk: ,23). Psychotropic drug use was also not different between groups. Furthermore, statistical models revealed no interaction between group and time in terms of NPI symptom severity or distress measures, indicating that dementia onset was not associated with worsening neuropsychiatric symptoms over time. Three incident dementia cases (11.5%) reverted from 'dementia' to 'MCI' status during the study; this is in line with previous studies showing relatively less stability of dementia diagnoses in individuals with premorbid cognitive challenges (39).

Limitations
Data for this study were drawn from the NACC data repository, which inherently limits the generalizability of ndings. AD is likely overrepresented in these data, and incident dementia cases within NACC may decline more signi cantly or at faster rates that incident dementia cases within the community. NACC included more than 35,000 participants enrolled as of the September 2016 data freeze, however only a small percentage of these had a diagnosis of BD, and ever fewer eventually developed dementia. Average follow-up length was also relatively short in some cases. As such, the relatively small sample of individuals retained for analysis may be viewed as a limitation of this study.
Another important limitation concerns the reliability and validity of self-reported BD diagnoses. A considerable number of participants were cognitively impaired, which may have led to inaccurate reports of medical and psychiatric histories. Results should be replicated in more carefully characterized samples. Furthermore, a number of potentially important variables were also not collected in NACC, including lifetime exposure to different pharmacological treatments, number of previous affective episodes and hospitalizations, all of which are known to affect cognition in BD (3,40). Age of BD onset is also not available in the NACC dataset. This information would be helpful to con rm whether earlyversus late-onset BD differ in their rates of cognitive impairment and dementia risk, as has been suggested previously (9).
Lastly, only a very small subsample of participants had available neuroimaging data, and these were obtained from a large collection of pooled scans from various centers with varying image acquisition protocols. Although the NACC protocol includes neuroimaging and cerebrospinal uid (CSF) biomarker data collection to support the etiological diagnosis of dementia, these data were unfortunately not available for participants used in this study. Thus, it must be acknowledged that that the etiological diagnosis of dementia was presumptive. Future work should characterize the structural and functional neurological changes that are associated with incident dementia in BD, as well as the presence and relevance of CSF biomarkers. Results from one recent study suggest that cognitively impaired BD patients do not display AD-speci c CSF biomarkers (41), however additional work is necessary to corroborate this nding and its association with cognition longitudinally.

Conclusions
Although epidemiologic evidence suggests that adults with BD may have increased dementia risk, to date, little is known about what neuropsychological changes may be harbingers of early neurodegeneration in these patients due to cognitive de cits being a core feature of BD independent of dementia. Our study provides preliminary evidence that some speci c cognitive de cits could be used as predictive measures for early diagnosis of dementia in patients with bipolar disorder. These ndings are of signi cant relevance given that the planet's population will include growing numbers of seniors, many of whom will have BD and may be in the early stages of dementia and need appropriate clinical management. Indeed, it has been estimated that one in four BD is over age 60 (42), and thus at high risk for neurodegenerative disease (9). Future studies may rely on the ndings provided in this study to further explore the issue of dementia in BD.