Bipolar disorder is a mental health condition characterised by recurrent episodes of depression and (hypo)mania. Recent findings suggest that people with bipolar disorder experience moderate cognitive impairments, 0.5–1 standard deviation (s.d.) below the normative mean, which are more pronounced during mood episodes but persist after the symptoms remit, during periods of euthymia.Reference Arts, Jabben, Krabbendam and van Os1,Reference Torres, Boudreau and Yatham2 Deficits appear across multiple domains, including processing speed, verbal memory and executive functioning.Reference Bourne, Aydemir, Balanza-Martinez, Bora, Brissos and Cavanagh3 Considerable evidence suggests that these impairments affect daily life functioning, possibly independent of mood symptoms.Reference Léda-Rêgo, Bezerra-Filho and Miranda-Scippa4 Findings indicate that cognitive difficulties are present in a proportion of patients with bipolar disorder,Reference Reichenberg, Harvey, Bowie, Mojtabai, Rabinowitz and Heaton5,Reference Burdick, Goldberg, Cornblatt, Keefe, Gopin and Derosse6 and more recently this heterogeneity has been characterised through the identification of different profiles of cognitive difficulties.
The evidence suggests that there are three or four discrete and coherent profiles, one cognitively intact and comparable to the general population, plus one or two subgroups presenting with selective moderate impairments, and a globally impaired subgroup with severe impairments across cognitive domains.Reference Green, Girshkin, Kremerskothen, Watkeys and Quidé7 Similar findings have been reported from studies with cross-diagnostic samples involving people with different diagnoses across the psychosis spectrum.Reference Van Rheenen, Lewandowski, Tan, Ospina, Ongur and Neill8,Reference Lewandowski, Sperry, Cohen and Ongur9
There is also a limited understanding of the factors underlying cognitive heterogeneity in bipolar disorder. An aspect of cognitive heterogeneity remaining unexplored is whether cognitive clusters represent different degrees of cognitive decline following illness onset and the extent of this putative decline across subgroups. A recent study with a cross-diagnostic sample defined cognitive decline as the discrepancy between current cognitive performance and premorbid IQ and reported evidence for a cluster characterised by a large cognitive decline, but limited changes for the other subgroups.Reference Karantonis, Rossell, Carruthers, Sumner, Hughes and Green10
Cognitive reserve might be a factor underlying this difference in cognitive decline between clusters.Reference Burdick, Russo, Frangou, Mahon, Braga and Shanahan11 It reflects resilience to brain pathology by minimising its effect on behavioural outcomes, such as symptom manifestation or cognitive dysfunction.Reference Stern12,Reference Stern13 Several proxy measures of cognitive reserve have been suggested, including years of education, occupational attainment and measures of vocabulary knowledge or reading abilities.Reference Sumowski, Chiaravalloti and DeLuca14 The concept of reserve has been extensively explored in neurological and psychiatric disorders such as dementia, multiple sclerosis and schizophrenia, representing individual differences in the capacity to compensate for age- and illness-related cognitive decline in the presence of neuropathology.Reference Sumowski, Chiaravalloti and DeLuca14–Reference Van Rheenen, Cropley, Fagerlund, Wannan, Bruggemann and Lenroot16 Previous research in bipolar disorder suggests an association of cognitive reserve with cognitive performance,Reference Grande, Sanchez-Moreno, Sole, Jimenez, Torrent and Bonnin17 but this has not been explored in the context of cognitive heterogeneity, particularly whether and to what extent cognitive differences between putative subgroups can be explained as a function of cognitive reserve.
Illness-history variables, such as the type and the number of previous mood episodes, have been investigated as putative underlying factors of cognitive heterogeneity. However, previous studies have reported inconsistent findings on the association of clinical characteristics and medication use with the profile of cognitive impairment.Reference Green, Girshkin, Kremerskothen, Watkeys and Quidé7 The inclusion of participants partially remitted or with subsyndromal symptoms may have resulted in these mixed findings between studies. Hence, it remains unclear whether cognitive performance across cognitive clusters is affected by patient differences in illness-related characteristics.
Here, we examine whether discrete cognitive profiles can be identified within a cohort of euthymic patients with bipolar disorder and test the hypothesis that cognitive reserve will differ between subgroups. We anticipate replicating the findings of previous cluster-analytic studies reporting on euthymic participantsReference Bora, Hidiroglu, Ozerdem, Kacar, Sarisoy and Civil Arslan18,Reference Roux, Raust, Cannavo, Aubin, Aouizerate and Azorin19 and we use independent cognitive measures to internally validate the identified clusters. Cognitive reserve and estimated postmorbid cognitive decline are examined as factors underlying potential differences in the cognitive course of putative subgroups. Differences in clinical characteristics are also examined among clusters. Parsing cognitive heterogeneity in bipolar disorder is important for delivering targeted interventions and potentially improving outcomes. A better understanding of the underlying factors associated with cognitive clustering may also inform intervention strategies, as well as help clinical services recognise earlier or more efficiently which patients require cognitive treatment.
Method
Study design
This is a cross-sectional secondary analysis of baseline data from the Cognitive Remediation in Bipolar (CRiB) study, a feasibility trial comparing cognitive remediation with treatment as usual in people with bipolar disorder.Reference Strawbridge, Fish, Halari, Hodsoll, Reeder and Macritchie20,Reference Strawbridge, Tsapekos, Hodsoll, Mantingh, Yalin and McCrone21 Written informed consent was obtained from all participants prior to inclusion. Baseline assessments were undertaken before random allocation to the treatment or the control group. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All procedures involving human patients were approved by the City Road & Hampstead NHS Research Ethics Committee (reference 15/LO/1557).
Participants
The sample comprised 80 out-patients with a DSM-5 diagnosis of bipolar disorder. All participants were fluent in English and aged between 18 and 65 years. Bipolar disorder subtype and eligibility were confirmed using the Mini International Neuropsychiatric Interview.Reference Sheehan, Lecrubier, Sheehan, Amorim, Janavs and Weiller22 Participants were free of acute mood symptoms for at least 1 month prior to inclusion. Remission at screening stage was defined using cut-off scores of ≤7 on the 17-item Hamilton Rating Scale for Depression (HRSD)Reference Hamilton23 and the Young Mania Rating Scale (YMRS).Reference Young, Biggs, Ziegler and Meyer24 Participants with a diagnosed neurological disorder or personality disorder and those misusing or dependent on alcohol or illicit substances over the previous 6 months were excluded.
Measures
Clinical assessment
A structured interview was used to collect information on demographic characteristics, illness-history variables and current medication use. Mood symptoms were assessed using the HRSD for depressive and the YMRS for hypomanic/manic symptoms.
Cognitive assessment
Clustering measures
Participants were administered an extensive cognitive battery in a standardised order. The battery included eight tests to be used as clustering tests: the Hotel test,Reference Manly, Hawkins, Evans, Woldt and Robertson25 Wechsler Memory Scale Fourth Edition (WMS-IV) Verbal Paired Associates immediate and delayed recall (VPA I and II),Reference Wechsler26 Wechsler Abbreviated Scale of Intelligence Second Edition (WASI-II) Matrix Reasoning subset,Reference Dumont, Willis, Veizel and Zibulsky27 Wechsler Adult Intelligence Scale Fourth Edition (WAIS-IV) Digit Span, Digit Symbol–Coding and Symbol Search,Reference Wechsler28 and the F-A-S letter verbal fluency test from the Delis–Kaplan Executive Function System.Reference Delis29
Non-clustering measures
Two measures of general cognitive performance were used for the internal validation of the emerging cognitive subgroups.Reference Karantonis, Rossell, Carruthers, Sumner, Hughes and Green10 Participants were administered the Montreal Cognitive Assessment (MoCA),Reference Nasreddine, Phillips, Bédirian, Charbonneau, Whitehead and Collin30 a compact screening instrument assessing multiple cognitive domains and giving a single score. We also used the cognitive subscale of the Functional Assessment Short Test (FAST),Reference Rosa, Sánchez-Moreno, Martínez-Aran, Salamero, Torrent and Reinares31 a clinician-rated measure examining cognitive functioning in the context of daily life situations and activities, with higher scores indicating greater difficulties.
IQ measures
Two intelligence measures were administered: an estimate of premorbid IQ, the Test of Premorbid Function (TOPF),Reference Wechsler32 and an estimate of verbal IQ, the Vocabulary subtest of the WASI-II.Reference Dumont, Willis, Veizel and Zibulsky27 Both represent cognitive reserve as measures of crystallised intelligence and have been shown to be resistant to the effects of ageing or illness-related decline.Reference Ryan, Sattler and Lopez33
Statistical analysis
Despite a lack of consensus on the required sample size to achieve adequate statistical power in cluster analysis, a minimum sample size of 2k (where k is the number of clustering variables) has been suggested, although the ideal sample size would be five times this number.Reference Dolnicar34 On the basis of this recommendation and our sample of 80 participants, we considered four clustering variables in this study.
All statistical analyses were conducted using the Statistical Package for Social Sciences (SPSS), version 25 for Windows, with a statistical significance of P < 0.05 (two-tailed) for all tests. The distributions were checked for normality across all measures using the Shapiro–Wilk test and log transformation was applied to conform non-normally distributed variables. Descriptive statistics were computed for all variables.
Cognitive domains for clustering
Raw scores of cognitive tests were transformed to demographically corrected standardised scores (z-scores; mean 0, s.d. = 1) based on the normative data for each test as provided by manuals. The Hotel test z-scores were inverted to be consistent with the direction of other measures, since for this test higher scores represent poorer performance. We computed z-scores for four cognitive domains: processing speed using the average scores of the Digit Symbol–Coding and the Symbol Search; working memory with the Digit Span forward, backward and sequencing scores; verbal learning and memory using the average scores of the VPA I and VPA II; and executive functioning using the average scores of the Hotel test, Matrix Reasoning and F-A-S letter verbal fluency test. A composite score of current global cognition was computed for each participant by averaging the z-scores of all the tests used for the clustering domains.
Identifying and validating cognitive clusters
The four cognitive domains were entered into a hierarchical cluster analysis to identify subgroups with homogeneous cognitive profiles according to their performance in these domains. Following Burdick et al,Reference Burdick, Russo, Frangou, Mahon, Braga and Shanahan11 we used Ward's linkage as the clustering method and squared Euclidean distance to estimate similarities between cases. To validate the initial clustering and evaluate the accuracy of participant allocation across clusters, we conducted a discriminant function analysis (DFA). The optimal number of clusters was determined on the basis of visual inspection of the dendrogram and the DFA coefficients. DFA examines the predictive power of different domain scores for every subgroup identified in the hierarchical cluster analysis and determines the probability of classification into a certain subgroup for every participant on the basis of these scores.
To estimate in which domains and to what extent these subgroups are distinct, cognitive profiles were compared to each other using multivariate analysis of covariance (MANCOVA) with least significant difference (LSD) for pairwise comparisons. Age was used as a covariate since age differences were evident between subgroups. For MANCOVAs with a significant main effect, post hoc comparisons between subgroups were corrected for multiple testing using a false discovery rate (FDR) of 5%.
Given that the cluster analysis was applied to differentiate participants on the basis of their cognitive performance, we anticipated significant between-group differences that would internally validate the cognitive clustering of the sample. Cohen's d effect sizes (defined as the mean group difference divided by the pooled s.d. and corrected for unequal group sizes) were calculated for significant post hoc comparisons. To further validate within the sample that the emerging cognitive profiles are truly distinct and not just artefacts of the measures used to identify them, subgroups were compared on non-clustering cognitive measures (MoCA and FAST cognitive subscale) using MANCOVA with age as a covariate and correcting for multiple comparisons with a 5% FDR.
Estimating cognitive decline and cognitive reserve
A proxy measure of postmorbid cognitive decline was estimated for each participant as the discrepancy between the current global cognitive performance (composite score) and the estimated premorbid IQ (TOPF score), following the methodology of previous studies.Reference Van Rheenen, Lewandowski, Tan, Ospina, Ongur and Neill8,Reference Karantonis, Rossell, Carruthers, Sumner, Hughes and Green10 Both the global cognition composite and the TOPF score were standardised to the normative performance of healthy controls according to the validation data.
Cognitive reserve was estimated for each participant through a composite score of three proxy measures: years of education and the two intelligence indices, the TOPF and the Vocabulary subtest of the WASI-II, according to previous recommendations.Reference Stern, Habeck, Moeller, Scarmeas, Anderson and Hilton15,Reference Van Rheenen, Cropley, Fagerlund, Wannan, Bruggemann and Lenroot16 Using a factor analysis, a single score was derived from these three variables for each participant.Reference Stern, Habeck, Moeller, Scarmeas, Anderson and Hilton15 This factor score accounted for 59.8% of the shared variance in the three variables and was used as a measure of cognitive reserve in all subsequent analyses. To ensure that cognitive reserve is an estimate independent of ageing and illness progression, this score was correlated to participants’ age, age at illness onset, illness duration and number of mood episodes.
Differences between subgroups in cognitive decline and cognitive reserve were examined using a MANCOVA (age as covariate) with an FDR of 5% to correct for multiple comparisons.
Comparisons on demographic and clinical characteristics
Subgroup comparisons were conducted to examine differences in non-cognitive variables using chi-squared (χ 2) tests or analyses of variance (ANOVAs) with Games–Howell post hoc correction for unequal variances and small sample sizes. Identified clusters were compared on demographic characteristics, clinical history, current mood symptoms, use of medication and previous service use. To estimate the magnitude of potential differences, Cohen's d was calculated using the same procedure. Cohen's w effect sizes were calculated for categorical variables.
Results
Sample characteristics
Table 1 provides details on the demographic and clinical characteristics of sample. There were no missing data for any clinical variables or cognitive tests. Overall, the sample performed mildly below the normative scores in two domains: verbal learning and memory (z = −0.26, s.d. = 1.1) and executive functioning (z = −0.22, s.d. = 0.66). Even smaller differences were detected in processing speed (z = −0.14, s.d. = 0.72), working memory (z = −0.08, s.d. = 0.61) and global cognition (z = −0.17, s.d. = 0.61).
Sample demographic and clinical characteristics (N = 80)
TOPF, Test of Premorbid Function; HRSD-17, 17-item Hamilton Rating Scale for Depression; YMRS, Young Mania Rating Scale.
Identifying cognitive clusters
Inspection of the dendrogram and the agglomeration coefficients suggested a four-cluster solution as the most appropriate fit for the sample: a subgroup of 25 participants (31%) remained cognitively intact, another subgroup of 15 participants (19%) had a selective deficit in verbal learning and memory, a third subgroup of 30 participants (37.5%) had intermediate impairment across all domains and a fourth subgroup of 10 participants (12.5%) had severe global impairments. The intact cluster performed slightly above the normative mean in three domains (by 0.2–0.3 s.d.) and 1 s.d. above the normative mean in verbal learning and memory. The second cluster showed a comparable performance to the intact subgroup in processing speed, working memory and executive functioning, but presented a selective impairment of 1 s.d. below the normative mean in verbal learning and memory. The intermediate cluster consistently performed 0.3–0.5 s.d. below the normative mean across all domains. The last cluster was characterised by severe impairments exceeding 1 s.d. below the norm in all domains. Figure 1 illustrates the cognitive profiles of the subgroups.
Cognitive profiles (domain mean and standard error) of the four subgroups (cognitively intact; selective deficit in verbal learning and memory; intermediate impairment across all domains; and severe global impairments). PrSp, processing speed; WM, working memory; VLM, verbal learning and memory; ExF, executive functioning; Composite, composite cognitive score.
The DFA revealed the presence of two significant discriminant functions, explaining 70% and 29% of the cluster membership variance respectively (Wilks’ λ = 0.11, χ 2 = 170.5, P < 0.001 and Wilks’ λ = 0.43, χ 2 = 63.2 P < 0.001 respectively). Verbal learning and memory were the main contributors to participant classification in function 1 (β = 0.7), followed by working memory in function 2 (β = 0.5). According to the DFA, 97.5% of the original grouped participants were correctly classified, suggesting a valid clustering of the sample. A scatterplot of participants per cluster is shown in Fig. 2.
Graphical agglomeration of the cognitive subgroups (cognitively intact; selective deficit in verbal learning and memory; intermediate impairment across all domains; and severe global impairments). Data plots represent participant scattering and clustering based on the discriminant function values. Group centroids represent mean scores for each cluster.
Subgroup cognitive profiles
Comparisons between subgroups on cognitive measures are presented in Table 2. As anticipated, the MANCOVA revealed a significant main effect of cluster on all clustering domains (all P < 0.001). Post hoc pairwise comparisons corrected for multiple testing (5% FDR) showed that participants in the severely globally impaired subgroup performed significantly lower in all clustering domains and the composite cognitive score compared with the cognitively intact subgroup (d = 2.4–4.1) and compared with the selectively impaired subgroup with the exception of verbal memory (d = 1.7–3.6). Smaller but still largely significant differences were found between participants with intermediate and severe global impairments in all domains but executive functioning (d = 1.3–3.1).
Comparison between subgroups on cognitive measuresa
LSD, least significant difference; Int, cognitively intact subgroup; IM, intermediate impairment subgroup; Sel, selective impairment subgroup; SG, severe global impairment subgroup; MoCA, Montreal Cognitive Assessment; FAST, Functional Assessment Short Test; TOPF, Test of Premorbid Functioning.
a. Means and standard deviations for all cognitive measures represent z-scores standardised to the normative data of each test. Age was included as a covariate in all analyses. Only post hoc comparisons significant at P > 0.05 after false discovery rate (5%) correction are reported. Cohen's d effect size corrected for unequal subgroup sample sizes (Hedges’ g).
Internally validating cognitive subgroups
The two non-clustering cognitive measures were significantly correlated with the four clustering domains across the sample. The MoCA showed moderate correlations with working memory (r = 0.29; P = 0.009) and processing speed (r = 0.37; P = 0.001) and moderate to large correlations with executive functioning (r = 0.42; P < 0.001) and verbal memory (r = 0.47; P < 0.001). The cognitive subscale of the FAST was moderately correlated with all domains (r = 0.27–0.37; all P < 0.01) but processing speed. The severe global impairment subgroup performed significantly worse than all other groups on the MoCA (d = 1.3–2.2), while both the intermediate impairment (d = 0.8) and the severe global impairment (d = 1.3) subgroups had poorer FAST scores than the intact subgroup.
Cognitive decline and cognitive reserve
The MANCOVA for the estimated cognitive decline and the cognitive reserve score revealed significant differences between subgroups for both measures (all P < 0.01). Details are reported in Table 2. The discrepancy between the current global cognition composite and the estimated premorbid cognitive functioning (TOPF score) indicated a decline for all subgroups, which gradually increased in magnitude from the cognitively intact to the severe impairment subgroup (Fig. 3). Compared with the intact participants, all other subgroups presented with significantly greater cognitive decline (d = 1.8–2.9). This difference was also significant between the selective impairment and the severe impairment subgroups (d = 1.3).
Estimated postmorbid cognitive decline per subgroup (cognitively intact; selective deficit in verbal learning and memory; intermediate impairment across all domains; and severe global impairments), calculated as the discrepancy between current global cognition composite score and premorbid cognitive functioning (Test of Premorbid Function) score. Significant differences (5% false discovery rate correction): *P < 0.1, **P < 0.001.
As hypothesised, the cognitive reserve score was not correlated with age (r = −0.09; P > 0.1), age at illness onset (r = −0.15; P > 0.1), illness duration (r = 0.08; P > 0.1) and number of mood episodes (r = −0.18; P > 0.1). The cognitively intact subgroup had a higher cognitive reserve compared with the intermediate impairment (d = 0.6) and the severe impairment (d = 1.2) subgroups, while a significant difference was found between the selective impairment and the severe impairment subgroups (d = 1). Cognitive reserve showed a small, non-significant correlation with estimated cognitive decline across the sample (r = −0.2; P = 0.08). However, examining this association within each group, cognitive reserve and cognitive decline were strongly correlated for the intact, the intermediate impairment and the severe impairment subgroups (r = −0.65 to −0.74; all P < 0.05), while the correlation was not significant for the selective impairment subgroup (r = 0.05).
Subgroup clinical characteristics
Table 3 reports the ANOVA results of cluster comparisons for demographic and clinical characteristics. Significant main effects of cluster were found for age, age at diagnosis, the number of psychological therapies previously attended and medication use (P = 0.02–0.04). From illness-history variables, only age at onset differed between subgroups, where the selectively and the intermediately impaired participants were diagnosed at an older age compared with the cognitively intact subgroup (d = 1 and d = 0.8 respectively). Intact participants had undertaken more psychological therapies than severely impaired participants (d = 1) and were taking fewer medications compared with the selectively and intermediately impaired subgroups (d = 0.8). A significantly smaller proportion of this subgroup used antipsychotics compared with all other subgroups (d = 0.7). No differences among subgroups were detected for any other clinical variables or mood measures.
Comparison between subgroups on demographic, clinical and illness-history variables
Int, cognitively intact subgroup; IM, intermediate impairment subgroup; Sel, selective impairment subgroup; SG, severe global impairment subgroup; HRSD-17, 17-item Hamilton Rating Scale for Depression; YMRS, Young Mania Rating Scale.
P values in bold indicate significance at P < 0.05.
a. Cohen's d effect size corrected for unequal subgroup sample sizes (Hedges’ g). For categorical variables, Cohen's w was computed based on the χ 2 statistic and reflects the mean difference between one subgroup and all the others.
Discussion
Using an extensive cognitive battery and hierarchical cluster analysis, this study examined the presence of discrete cognitive profiles in a cohort of euthymic people with bipolar disorder to evaluate a clustering solution independent of mood symptoms. We examined the characteristics of these clusters, specifically related to cognitive reserve. Approximately one-third of participants appeared cognitively intact (31%) and another third demonstrated intermediate deficits across all domains (37.5%), with the remainder presenting a selective impairment in verbal learning and memory (19%) or showing severe deficits across all domains (12.5%). The most discriminating cognitive domains for the profiles were verbal learning and memory, followed by working memory and processing speed. Cognitively impaired subgroups presented with greater cognitive decline and poorer cognitive reserve, but few differences were detected in clinical characteristics.
How do the identified cognitive profiles relate to previous findings?
The four clusters found in this study are in line with previous work on euthymic individuals with bipolar disorder,Reference Bora, Hidiroglu, Ozerdem, Kacar, Sarisoy and Civil Arslan18,Reference Roux, Raust, Cannavo, Aubin, Aouizerate and Azorin19 as well as cross-diagnostic studies including participants from the wider schizophrenia–bipolar disorder spectrum.Reference Lewandowski, Sperry, Cohen and Ongur9 The proportion of cognitively intact participants (31%) is very similar to reports in other studies (20–45%),Reference Green, Girshkin, Kremerskothen, Watkeys and Quidé7 and above average performance in verbal learning and memory in this subgroup has also been reported.Reference Roux, Raust, Cannavo, Aubin, Aouizerate and Azorin19 A relatively small proportion (12.5%) presented with severe deficits across all domains, previously reported to comprise 10–35% of people with bipolar disorder.Reference Green, Girshkin, Kremerskothen, Watkeys and Quidé7 A possible explanation for the proportion being at the lower end of this range is that our sample included only individuals in full remission, with higher overall premorbid cognitive functioning and more years in education than previous studies, which might indicate a high-performing sample compared with other cohorts (supplementary Table 1, available at https://doi.org/10.1192/bjo.2020.111). The other two clusters included participants who either experienced a significant impairment in a single domain but otherwise remained intact (19%), or participants with intermediate impairment across all domains (37.5%). These clusters correspond to subgroups described as having ‘selective’ or ‘moderate’ impairment in previous studies reporting four clusters (15–40%).Reference Green, Girshkin, Kremerskothen, Watkeys and Quidé7
The division of moderately impaired participants into two distinct clusters was the main difference from previous studies reporting three-cluster solutions. One factor affecting potential clustering is the definition of euthymia used for inclusion. In this study we included only individuals in full remission and previous studies with similar definitions also found four-cluster solutions, with subgroups impaired in a single domain or an intermediate impairment across all domains.Reference Bora, Hidiroglu, Ozerdem, Kacar, Sarisoy and Civil Arslan18,Reference Roux, Raust, Cannavo, Aubin, Aouizerate and Azorin19 In contrast, studies with less restrained definitions or including participants in partial remission have mostly reported three-cluster solutions.Reference Burdick, Russo, Frangou, Mahon, Braga and Shanahan11,Reference Jensen, Knorr, Vinberg, Kessing and Miskowiak35,Reference Russo, Van Rheenen, Shanahan, Mahon, Perez-Rodriguez and Cuesta-Diaz36 (supplementary Table 1). Although there is no clear explanation for this disparity, it is possible that reduced cognitive performance in otherwise intact domains due to residual depressive symptoms leads to a shared cluster between participants with selective and intermediate deficits.
The performance of each subgroup across the clustering domains suggested largely distinct cognitive profiles, apart from the cognitively intact and the selective impairment subgroups, where the difference was significant only for verbal memory. This distinction was partially validated by the non-clustering measures. Although the clustering and non-clustering cognitive measures showed small to moderate correlations between subgroups, significant subgroup differences validating the identified profiles for this sample were only found between the intact and the severely impaired participants. These two subgroups were clearly separated on both non-clustering measures, whereas differences observed between the selective impairment and intermediate impairment subgroups were minimal.
Despite multiple studies supporting the existence of distinct cognitive profiles among people with bipolar disorder, these results were solely based on behavioural measures analysis, while potential biological underpinnings of cognitive variability remain largely unknown. Recent neuroimaging findings have pointed to weaker interregional connectivity as a neural mechanism possibly underlying cognitive heterogeneity in subgroups of people with bipolar I disorder clustered on the basis of their performance on executive functioning tasks.Reference Kollmann, Yuen, Scholz and Wessa37 Although this study differed from typical cluster-analytic designs in terms of grouping participants according to their response pattern (i.e. encompassing strengths and deficits in their performance) rather than their normative cognitive performance, it did provide initial evidence on the neurological background of distinct cognitive profiles in bipolar disorder.
What is the role of cognitive reserve?
Reduced premorbid IQ has been associated with the severe impairment subgroup in previous bipolar disorder and cross-diagnostic cluster-analytic studies.Reference Karantonis, Rossell, Carruthers, Sumner, Hughes and Green10,Reference Burdick, Russo, Frangou, Mahon, Braga and Shanahan11 Our study cultivates these findings and indicates a linear pattern of increasingly reduced cognitive reserve across subgroups, gradually reducing from the intact to the severe global impairment subgroup. The reverse pattern was observed for the estimated cognitive decline, and the association between cognitive reserve and cognitive decline was significant for all participants but the selectively impaired ones. On the basis of these findings, it is plausible that differential cognitive performance across subgroups is a function of cognitive reserve for the intact, the intermediate impairment and the severe impairment subgroups. This is consistent with the theoretical concept of cognitive reserve in dementia and schizophrenia spectrum or mood disorders, as a protective mechanism against cognitive decline caused by neuropathology itselfReference Stern13,Reference Watson and Joyce38 or against the adverse cognitive effects of certain treatments (e.g. memory performance after electroconvulsive therapy).Reference Legendre, Stern, Solomon, Furman and Smith39
Although cognitive decline after symptom onset was evident across all subgroups, the severely impaired subgroup showed the greatest decline and had the poorest cognitive reserve. Cognitive reserve potentially modifies cognitive performance and explains individual differences in the cognitive course of people with bipolar disorder, with worse outcomes for those with poorer reserve.Reference Grande, Sanchez-Moreno, Sole, Jimenez, Torrent and Bonnin17,Reference Anaya, Torrent, Caballero, Vieta, del Mar Bonnin and Ayuso-Mateos40 Recent evidence points to a subgroup of people with bipolar disorder sharing genetic and symptomatic characteristics with schizophrenia.Reference Bora and Pantelis41,Reference Service, Vargas Upegui, Castaño Ramírez, Port, Moore and Munoz Umanes42 This subgroup presents with a cognitive profile comparable to that of people with schizophrenia, including a neurodevelopmental trajectory characterised by poor cognitive reserve.Reference Bora43,Reference Van Rheenen, Lewandowski, Bauer, Kapczinski, Miskowiak and Burdick44 This neurodevelopmental hypothesis is further supported by evidence for the existence of discrete cognitive clusters in young offspring of people with bipolar disorder prior to illness onset.Reference Bora, Can, Ildız, Ulas, Ongun and Inal45
In our sample, participants with severe deficits across domains potentially represent this subgroup of people with bipolar disorder following a cognitive course defined by reduced cognitive reserve. This argument is strengthened by non-significant or inconsistent findings for subgroup differences in clinical and illness-history characteristics (e.g. diagnostic subtype, illness duration, number of episodes), suggesting that cognitive variability in people with bipolar disorder cannot be entirely attributed to illness progression. Pending confirmation by longitudinal studies, cognitive reserve may be a factor driving disparate illness trajectories and a characteristic that may predict the extent of cognitive impairment in people with bipolar disorder.
Do clusters differ in their clinical characteristics?
No significant differences were observed between subgroups in demographic characteristics and most illness-history variables, which suggests that subgroup allocation and cognitive variability in our sample was largely independent of clinical characteristics. Likewise, subthreshold depressive and (hypo)manic symptoms were balanced between subgroups, which has been previously reported for cluster-analytic studies in bipolar disorder.Reference Green, Girshkin, Kremerskothen, Watkeys and Quidé7
Although the duration of illness and the number of mood episodes did not differ between subgroups, cognitively intact participants were diagnosed at a younger age than participants in the selective impairment and intermediate impairment subgroups. People with bipolar disorder experience significant delays in receiving the correct diagnosis and in receiving treatment.Reference Dagani, Signorini, Nielssen, Bani, Pastore and Girolamo46 As a result, differences in age at diagnosis may not reflect differences in age at symptom onset. To examine that further, we computed the duration of untreated illness (DUI), defined as the discrepancy between the time when they first experienced symptoms and when they received their diagnosis, and compared it across subgroups. Although cognitively intact participants received their diagnosis 7 years on average after symptom onset and DUI exceeded 10 years for all other subgroups, this difference was not statistically significant. Thus, our findings do not support that greater DUI is associated with more severe illness course in terms of cognitive impairments.Reference Drancourt, Etain, Lajnef, Henry, Raust and Cochet47
Subgroup differences were evident in treatment characteristics. To the best of our knowledge, this is the first study reporting that cognitively intact participants had undertaken more psychological therapies than those with severe impairment, but it is unclear whether this reflects a neuroprotective effect of psychological therapies or if it simply reflects the capacity of participants in this subgroup for better treatment access and adherence. A smaller proportion of intact participants were using antipsychotics compared with the other subgroups. This was despite the lack of any significant subgroup differences in history of psychosis and has been previously reported in cluster-analytic studies.Reference Bora, Hidiroglu, Ozerdem, Kacar, Sarisoy and Civil Arslan18 The relationship between treatment and potential cognitive outcomes is complex, as greater use of antipsychotics may be an indicator of greater illness severity, particularly regarding increased risk for manic relapse. Studies examining the cognitive course of patients who discontinued or reduced their antipsychotic medication may clarify this question.Reference Albert, Randers, Allott, Jensen, Melau and Hjorthøj48
Limitations
This was a cross-sectional study so we cannot speculate on the stability of these cognitive subgroups. Likewise, estimated cognitive decline following illness onset was calculated using cross-sectionally collected measures. In addition, the sample size was relatively small, although it did meet the minimum requirements for a cluster analysis. Therefore, findings on the role of cognitive reserve need to be interpreted with caution and require further replication in larger studies. There is a possibility that effect sizes were inflated owing to our sample size, but we did use a conservative approach for multiple comparison correction to reduce false positives. We therefore consider the identified subgroups and the detected differences reasonable. The neuropsychological battery used was less extensive, with fewer tests per domain compared with some of the previous studies, but cognitive scores did not differ significantly within each domain and the variation in participant scores from the domain mean was small, indicating limited estimation bias. Finally, we did not measure psychiatric comorbidities that could potentially affect cognitive performance.
What are the clinical implications?
We recently reviewed the literature on cognitive remediation therapies targeting cognitive and functional outcomes for people with bipolar disorder.Reference Tsapekos, Seccomandi, Mantingh, Cella, Wykes and Young49 Findings are promising but still inconsistent across studies. Cognitive heterogeneity in study samples may underly this inconsistency. Individuals with different cognitive profiles can have different outcomes and some adaptations may be necessary in treatment provision. Different patient clusters may require modifications to adhere to and benefit from treatment, for example longer or more intensive therapy periods for those more severely impaired.Reference Tsapekos, Strawbridge, Cella, Wykes and Young50 Our findings also suggest that considering the concept of cognitive reserve might be relevant in the context of designing and delivering cognitive interventions, either as a potential treatment target or as a factor enabling treatment engagement and outcomes. Clinical services for people with bipolar disorder should introduce cognitive assessment into their screening process for new or recurring patients. The cognitive domain mostly contributing to differentiation of the four cognitive profiles was verbal learning and memory, and previous research has associated verbal memory performance with future functional outcomes.Reference de Mar Bonnin, Gonzalez-Pinto, Sole, Reinares, Gonzalez-Ortega and Alberich51 Hence, it might be a suitable domain for clinical services to screen in order to quickly and effectively differentiate between cognitively intact and compromised patients.
Supplementary material
Supplementary material is available online at https://doi.org/10.1192/bjo.2020.111.
Data availability
The anonymised data-set is available from the corresponding author on request.
Acknowledgements
We are grateful to all CRiB study participants and team members.
Author contributions
D.T. and A.H.Y. conceived and planned this study. D.T., R.S. and T.M. were involved in data collection. D.T. carried out the analysis and prepared the first manuscript, with support from R.S. and T.M.. M.C., T.W. and A.H.Y. supervised the project. All authors provided feedback and contributed to the final manuscript.
Funding
This paper represents independent research part-funded by the National Institute for Health Research (NIHR) Maudsley Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King's College London. The views expressed are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care. D.T. would like to acknowledge that this paper was supported by the Onassis Foundation (scholarship: F ZO077-1/2018-2019). A.H.Y. and T.W. would like to acknowledge their NIHR Senior Investigator awards.
Declaration of interest
A.H.Y. reports paid lectures and advisory board membership for the following companies with drugs used in affective and related disorders: AstraZeneca, Eli Lilly, Lundbeck, Sunovion, Servier, LivaNova, Allegan, Bionomics, Sumitomo, Dainippon Pharma and Janssen; consultancy to Johnson & Johnson and LivaNova; principal investigator on the Restore-Life VNS registry study funded by LivaNova and on ESKETINTRD3004 funded by Janssen. R.S. reports a paid lecture from Lundbeck.
ICMJE forms are in the supplementary material, available online at http://dx.doi.org/10.1192/bjo.2020.111.
Open access
eLetters
No eLetters have been published for this article.