Impaired Retrieval of Semantic Information in Bipolar Disorder: A Clustering Analysis of Category-Fluency Productions

Abstract

On category-cued verbal fluency tasks, such as animal naming, respondents often report exemplars in semantically related clusters. We (Sung et al., 2012) used this tendency to elucidate sources of semantic dysfunction in adults with schizophrenia (SZ). Many patients with bipolar disorder (BD) show cognitive deficits that are similar to but milder than those seen in SZ. Whether this similarity extends to the functioning of the semantic system is unclear. To test the hypothesis that it does, we adapted a clustering technique called singular value decomposition (SVD) to investigate the clustering pattern of semantic retrieval in BD. Two category-fluency tasks (animal and supermarket-item naming) were administered to 98 adult outpatients with BD and 98 healthy adults (NC) who matched the BD group in age, sex, education, and estimated premorbid IQ. Results of clustering analysis showed that patients with BD produced less coherent category clusters than healthy adults. Specifically, patients with BD showed less coherent clusters of low-frequency animal names, but their overall productivity was not more impaired than the NCs’. In the supermarket condition, patients not only showed incoherent clustering of named supermarket items regardless of their frequencies, but also produced smaller numbers of exemplars than NCs did. The semantic system abnormalities shown by adults with BD were similar to those we found previously in persons with SZ, although the group differences were smaller. Overall, these results point to a concept retrieval/access deficit in BD and underscore the importance of analyzing the content of category-fluency productions.

Bipolar disorder (BD) is conceptualized primarily as involving dysregulation of mood rather than disordered thinking, which is a hallmark of schizophrenia (SZ). However, a growing body of evidence suggests that persons with BD can have trait-like impairments of attention, memory, executive functioning, and other cognitive abilities (Bearden, Hoffman, & Cannon, 2001; Bilder, 2001; Dickerson et al., 2004; Quraishi & Frangou, 2002; Schretlen et al., 2007; Seidman et al., 2002). These studies generally show that patients with BD and SZ demonstrate similar qualitative deficits that differ mainly in severity.

Among cognitive domains that have been studied, the semantic system, which comprises semantic knowledge and its functions (e.g., semantic activation/inhibition and spreading activation), has been assessed through category-cued fluency tasks. Typically, category-fluency comparisons of patient and healthy control (NC) groups are based on the number of acceptable category exemplars named, that is, overall productivity. In these studies, patients with BD often gave fewer category exemplars than the healthy adults did, but more than patients with SZ (Dickerson et al., 2004; Lebowitz, Shear, Steed, & Strakowski, 2001; Schretlen et al., 2007; Torrent et al., 2007). However, this finding is far from uniform (Jamrozinski, Gruber, Kemmer, Falkai, & Scherk, 2009; Rossell, 2006). Regardless of the findings, these studies are limited by their focus on overall productivity. They provide no information about the possible sources of semantic system dysfunction, and they likely will fail to capture deficits that are not closely linked to productivity.

Individual and group differences in productivity on category-fluency tasks may or may not reflect impairments of the underlying semantic system. For example, reduced category fluency in SZ has been attributed variously to a literal degradation of semantic information (i.e., a lack of concepts, or storage deficit) or impaired retrieval of concepts from an otherwise intact semantic system (i.e., atypical semantic activation/inhibition; functional deficit; Aloia, Gourovitch, Weinberger, & Goldberg, 1996; Joyce, Collinson, & Crichton, 1996; Rossell, Rabe-Hesketh, Shapleske, & David, 1999; Sung et al., 2012). Decreased productivity does not exclusively support one of these explanations over the other, because both can accommodate it equally well. The same applies to category fluency in BD if impaired productivity is observed. Conversely, normal productivity in BD does not necessarily mean the semantic system is intact. Although either deficit could lead to reduced productivity, impaired semantic activation or related functions are probably more closely related to the content of exemplars reported and the extent to which semantically related items are clustered during retrieval. Thus, although impaired productivity likely denotes some impairment of semantic knowledge, the reverse is not necessarily true because some semantic deficits might not reduce productivity. Consequently a content analysis (e.g., Sung et al., 2012; Troyer, Moscovitch, & Winocur, 1997; Van Beilen et al., 2004) of verbal fluency data becomes critical.

One type of content analysis that has been applied to category-fluency productions involves generating a visual representation of the semantic clusters of exemplars named by patients and comparing them with those produced by healthy adults. Multidimensional scaling (MDS) is the most frequently used technique for this (e.g., Aloia et al., 1996; Paulsen et al., 1996; Rossell et al., 1999; Sumiyoshi et al., 2005). This approach is based on the observation that, for a given category cue such as “animals,” respondents typically begin naming a cluster of semantically related words, such as “wild/African animals,” and then switch to a different cluster, such as “domestic/farm animals” (Sung et al., 2012; Troyer et al., 1997). The clustering of semantically related words is thought to depend on automatic spreading activation and the organization of concepts, both of which are well-recognized roles of the temporal lobe (Moscovitch, 1994; Schretlen et al., 2007; Troyer, Moscovitch, Winocur, Alexander, & Stuss, 1998).

To our knowledge, however, only one recent clustering analysis of category fluency in patients with BD has been reported (Chang et al., 2011). It was based on a Korean sample, and the investigators found that persons with BD showed less coherent clustering of category exemplars than healthy adults. Whether the results of obtained by Chang et al. would generalize to an English-speaking sample of persons with BD is unclear. In any case, the obtained results also require confirmation because Chang et al., 2011 used MDS to analyze their data, and MDS has significant limitations when applied to category-fluency data, as we discuss below.

The current study was designed to achieve three goals. First, we aimed to compare detailed semantic clustering patterns in the category examples generated by outpatients with BD and matched NCs based on a clustering procedure (singular value decomposition, or SVD) that overcomes some of the limitations of MDS when applied to category-fluency data. Second, we wanted to test two possible competing explanations (storage vs. functional deficit) of impaired semantic fluency in BD based on clustering patterns, which cannot be tested by examining productivity. Third, we aimed to determine whether the source of semantic dysfunction in BD could be the same as that seen in SZ. Achieving these goals would show that decreased productivity is not an obligatory manifestation of semantic impairment in persons with BD.

The SVD procedure we adapted for the current study is different from other methods used in clustering analyses of verbal fluency data, such as MDS and hierarchical clustering (e.g., Chang et al., 2011; Rossell et al., 1999; Sumiyoshi et al., 2005). Briefly, SVD is a general matrix-factorization technique in which eigenvalue decomposition (the mathematical basis for factor analysis) is a special case. The matrix to be analyzed in this study consists of words generated on category-fluency tasks. A typical result of SVD is the grouping of similar word vectors that can be used to approximate the original matrix with a dimensionality less than that of the original input matrix. SVD has been applied to many areas of science as a clustering technique (Alter, Brown, & Botstein, 2000; Frieze, Kannan, & Vempala, 1998; Landauer, 2007), including our recent application of it to category fluency data produced by adult outpatients with SZ (Sung et al., 2012). One advantage of SVD over MDS when applied to fluency data is that it allows one to analyze almost the entire dataset to examine more detailed patterns of semantic clusters. This overcomes the critical limitation of MDS and other clustering techniques, which is that these techniques are limited to examining only 10–20 most frequently named words (e.g., Chang et al., 2011; Sumiyoshi et al., 2005). Because any differences between BD and NC groups are expected to be subtle, SVD may be more suitable than other methods for this investigation. (See Sung et al., 2012, for a detailed discussion on SVD applied to fluency data).

Method

Participants

Participants with BD were recruited from Johns Hopkins-affiliated outpatient clinics, a psychiatric day hospital, and via flyers posted at Johns Hopkins Hospital. The patient sample consisted of 127 adults who met Diagnostic and Statistical Manual of Mental Disorders (4th ed.; DSM–IV) criteria for bipolar disorder (American Psychiatric Association, 1994). Diagnoses were made by either a psychiatrist or a neuropsychologist using the Diagnostic Interview for Genetic Studies (Nurnberger et al., 1994) or Mini-International Neuropsychiatric Interview (Sheehan et al., 1998). Of the 127 patients, eight were tested twice because they served in two studies, and we dropped their second protocols. Another 21 patients were excluded based on a history of substance dependence in the past 12 months, intellectual disability, dementia, stroke, or traumatic brain injury with more than one hour loss of consciousness. This left 98 adult outpatients with either BD I (88%) or BD II (12%) disorder. The mean age at diagnosis was 24.0 (SD = 8.5) years. Of 94 patients, 77 (82%) had been hospitalized at least once for their illness (data were missing for 4 patients). Most (90%) were tested a year or more after their most recent hospitalization. Prior suicide attempts were reported by 53% of participants for whom such information was available (n = 93). Of the 83 participants for whom such information was available, the most recently experienced episode involved depression for 47%, mania for 40%, and a mixed presentation for 13%. At the time of study, 40% of participants were prescribed Lithium, 45% were prescribed antipsychotics, 52% were prescribed antidepressants, and 47% were prescribed anticonvulsants. The patient sample included slightly more women (61%) than men (see Table 1). Participants ranged from 20 to 75 years of age and completed an average of 13.9 years of education (SD = 3.2).

Table 1.

Demographics of Participant Samples

Characteristic	Bipolar disorder (n = 98)	Healthy adults (n = 98)
Bipolar I/II: number (%)	86 (87.8%)/12 (12.2%)	N/A
Sex: female %*	61.2%	65.3%
Race: White/Black/other %**	54.1/39.8/6.1%	75.5/21.4%/3.1%
Age: mean (SD) years*	40.7 (10.7)	41.2 (11.1)
Education: mean (SD) years*	13.9 (3.2)	14.1 (2.8)
Estimated IQ:1 mean (SD)*	102.3 (11.4)	102.3 (10.8)

¹Based on Hopkins Adult Reading Test (Schretlen, et al., 2009).

^*ns.

^**χ² = 7.8, p < .05.

A group of healthy adults was drawn from the Aging, Brain Imaging, and Cognition (ABC) study of normal aging (Schretlen, Testa, Winicki, Pearlson, & Gordon, 2008). The 394 community-dwelling ABC study participants were recruited via random digit dialing or by calling randomly selected listings from the residential telephone directories for the Baltimore, Maryland and Hartford, Connecticut metropolitan areas. In addition to the exclusion criteria listed for BD participants, NCs were excluded from the ABC study if they had BD, schizophrenia, current major depression or substance abuse/dependence, any other medical or neurological condition commonly associated with cognitive impairment, or a score below 24/30 on the Mini-Mental State Exam (Folstein, Folstein, & McHugh, 1975). This resulted in 327 ABC participants with usable data. For each of the 98 participants in the BD group, we then selected from the ABC sample a sex-matched, age-matched (±5 years), and estimated premorbid IQ-matched sample of healthy adults. Premorbid IQ estimates were based on the Hopkins Adult Reading IQ (Schretlen et al., 2009). In four cases, a healthy woman was paired with a BD man to maintain the other matching criteria. Both groups included slightly more women than men. Participants in the NC group ranged from 20 to 75 years of age (M = 41.2, SD = 11.1) and had completed an average of 14.1 year of education (SD = 2.8). As shown in Table 1, the groups did not differ significantly in age, sex, years of education, or estimated premorbid IQ (ps > 0.05), but the BD group included more African Americans than the NC group. The Johns Hopkins Medicine Institutional Review Board approved the current study and the studies from which subjects were drawn, and each person gave written informed consent to participate.

Procedure

All participants completed two category-cued fluency tasks (animal names and supermarket items) from the Calibrated Ideational Fluency Assessment (CIFA; Schretlen & Vannorsdall, 2010) as part of a larger neuropsychological assessment. For each cue, participants were asked to verbally produce as many category examples as possible in 60 s. Their verbal responses were recorded verbatim by the examiner and later transferred into electronic form.

Analysis

After removing rule breaks (repeated words and noncategory examples) from the raw data, we converted plurals to singulars on animal naming and singulars to plurals on supermarket item naming for consistency in grammatical number and convenience. There were 64 (2.4%) rule breaks among the 2,676 animals named by all NCs and 63 (2.6%) rule breaks among the 2,392 animals named by patients, an insignificant difference (χ²_{1, N = 196} = 0.30, p = .58). There were 46 rule breaks among the 2,676 supermarket items named by NCs and 59 rule breaks among the 2,392 items named by patients. Although BD patients seemed to make more rule breaks in the supermarket condition than NCs, a χ² test indicates that the frequencies of rule breaks were only marginally different in two groups (χ²_{1, N = 196} = 3.48, p = .06).

A 2 × 2 repeated-measure ANOVA was performed, with subject group (BD, NC) as the between-group factor and fluency category (animals, supermarket items) as the within-group factor on the correct words reported. For cluster analysis, one word-by-subject matrix for each category condition was constructed for each subject group. Thus, four separate matrices were analyzed. The procedure for SVD analysis is virtually identical to that used in our previous study of persons with SZ (Sung et al., 2012). The matrix of animal names reported by healthy adults consists of 247 rows (unique animals named) and 98 columns (one for each participant), yielding a 247 × 98 input matrix. Each cell, c_ij, of the matrix has value of 1 if the jth subject said the ith animal name or 0 if not. We extracted word vectors up to the 25th dimension using the PROPACK software for SVD (Larsen, 2004) for Matlab (Version R2011a; Mathworks, Natick, MA). We assumed that the meaningful number of dimensions would be far fewer than 25. The three remaining matrices were constructed and analyzed in the same way. The BD group reported 265 different animals for a 265 × 98 matrix. The NC and BD groups reported 504 and 496 different supermarket items, from which we constructed input matrices of 504 × 98 and 496 × 98, respectively, for SVD analyses.

Results

Analysis of Variance

A repeated measure ANOVA of correctly named exemplars showed a large main effect for semantic category, F(1, 194) = 146.8, p < .01, partial ${\mathrm{\eta }}_p^2=0.43$ and a significant but weaker main effect for diagnostic group, F(1, 194) = 6.9, p < .01, ${\mathrm{\eta }}_p^2=0.03$. The effect size of 0.03 for the group difference is very small, according to Cohen (1992). The two-way interaction between group and task was not significant, F(1, 194) = 2.6, p < .11, ${\mathrm{\eta }}_p^2=0.01$. Post hoc comparisons showed that the main effect of group on animal naming was not significant, F(1, 194) = 3.4, p < .07, ${\mathrm{\eta }}_p^2=0.02$, whereas the main effect of group on supermarket item naming was F(1, 194) = 7.9, p < .01, ${\mathrm{\eta }}_p^2=0.04$. Nevertheless, the effect sizes for of the group differences were very small for both semantic category comparisons.

Clustering Analysis

Animal category: 2-D cluster representations.

Clustering patterns for the top 40 animals named by each group are presented in 2-D space in Figure 1. Animals with Frequency Ranks 1 through 20 are shown in Figure 1A and 1B for NC and BD, respectively. Those with Ranks 21 through 40 are shown in Figure 1C and 1D for NC and BD to avoid clutter. Note that all plots are on the same scale. Although all words were analyzed up to 25th dimensional solutions through SVD, we find that comparing the top 40 words in terms of vector dimensions 2 and 3 tends to be most informative. The degree of clustering between any two exemplars in all figures is determined by the angle between the vectors representing each exemplar, not by the Euclidean distance between each work, as is the case in MDS. The angle shows a better clustering relationship between words than other measures, such as Euclidean distance, in SVD (Landauer, McNamara, Dennis, & Kintsch, 2007).

Figure 1.

Clustering of top 40 animal names in 2-D vector space (Dimensions 2 and 3) are presented in four panels. (A) Animals with ranks of 1 through 20 named by NCs (1–10 in black and 11–20 in gray; see Appendix for the full list of ranks). The actual but hidden vector for each animal is a directional line from the origin of axes to the starting letter of an animal name (e.g., “h” of horse). (B) Animals (ranks 1–20) named by patients with BD. (C) Animals with ranks of 21–40 named by NCs (21–30 in black and 31–40 in gray). (D) Animals (ranks 21–40) named by patients with BD.

Figure 1A and 1C generally show that the NC group produced three major clusters of animals. These appear to comprise domestic/farm animals (e.g., cow, pig, horse, chicken, sheep, goat, etc.), wild/African animals (e.g., monkey, lion, tiger, giraffe, elephant, etc.) and a third category of other animals (e.g., fish, bird, mouse, whale, alligator, crocodile, shark, snake, etc.). This group of animals reported by healthy adults can be further divided into subgroups of reptiles and sea animals when 4th and 5th dimensions are added (see next section). Finally, and perhaps counterintuitively, the two top-ranked animals (dog and cat in Figure 1A) are not clearly associated with other domestic/farm animals.

Patients with BD showed some animal clusters in Figure 1B and 1D, but they were less tidy than those shown by NCs. Again, the two top-ranked animals (dog and cat) were not clustered with other domestic animals in Figure 1B. Animals belonging to the wild/African cluster were widely separated from each other (e.g., elephant and hippopotamus) and overlapped many domestic animals (e.g., pig and donkey) in the analysis of exemplars reported by the BD group. Also, animals such as dolphin, shark, whale, alligator, and crocodile were not separated into subclusters of reptiles and sea animals (as was the case for the NC group) when a 4th dimension was added (see below).

Animals: cosines of vector angles.

Figure 2 shows the cosines of angles between each of three selected animals and the 40 top-ranked animals in three different dimensional solutions (1st–3rd, 1st–4th, and 1st–5th). The cosine values of 1, 0, or −1 indicate that the directions of any two word vectors are identical (highly associative), orthogonal (independent), or opposed (exclusive), respectively, in multidimensional space. The numbers on the x axis indicate the ranks of the top 40 animals, whose names are presented in the figure caption (also see the Appendix). The rank order for each animal name is presented in a parenthesis in this section.

Figure 2.

Cosines of selected word vectors against top 40 animals (A) and (B). Cosines of angles between cat (1) and the top 40 animals by NC and BD, respectively (C) and (D). Cosines of vector angles between elephant (5) and top 40 animals by NC and BD. (E) and (F) cosines of vector angles between whale (27) and top 40 animals by NC and BD. The horizontal axis represents top 40 animals. They are, numbering by 5, cat (1), dog, lion, tiger, elephant (5), giraffe, bear, horse, zebra, monkey (10), snake, cow, bird, pig, deer (15), fish, mouse, rabbit, hippopotamus, rhinoceros (20), rat, alligator, squirrel, gorilla, sheep (25), chicken, whale, goat, leopard, eagle (30), crocodile, fox, kangaroo, shark, lizard (35), raccoon, ape, dolphin, duck, and donkey (40).

When cat (Rank 1) and other top 40 animals are compared in Figure 2A for NC, we see that cat (1) does not show signs of clustering (e.g., no clear ups and downs) with other domestic animals, for example, horse (8), cow (12), and pig (14). This echoes what was already seen in Figure 1A. Other top-ranked animals, e.g., dog (2), lion (3), and tiger (4) also show no clear pattern of clustering with other semantically associated animals. Although we saw that lion (3) and tiger (4) are clearly clustered with other wild/African animals in Figure 1A, when vector spaces with higher dimensionality (>2) are examined as in Figure 2A, they do not seem to form clear clusters with other wild/African animals. This absence of apparent clustering of top ranked animals was also seen in our previous study and some of other non-SVD studies (Aloia et al., 1996; Sumiyoshi et al., 2005; Sung et al., 2012).

We see the same patterns of cosine values for cat (1) with the other 40 animals in Figure 2B for persons with BD. Here, cat (1) shows a stronger association with dog (2) than with lion (3) or tiger (4) than it did in the NC data, but this is not very remarkable given the representativeness of these animals in their respective clusters. In fact, cat (1) is more closely associated with lion (3) and tiger (4) than with horse (8) and cow (12) in BD, which is also true in NC.

When elephant (5) is compared with top 40 animals in NC (Figure 2C), we see that it is associated strongly with wild/African animals such as giraffe (6), bear (7), zebra (9), monkey (10), hippopotamus (19), and so forth, and more weakly with other animals such as horse (8), cow (12), and pig (14). Elephant does not show strong association with different types of wild animals such as alligator (22), whale (27), eagle (30), crocodile (31) or shark (34), which form separate clusters, as seen in Figure 1A and 1C.

In BD (Figure 2D), patterns of associations between elephant (5) and the 40 top-ranked animals are mixed in that some have cosine values similar to those seen in the NC group, but others do not. Elephant (5) is less strongly associated with bear (7), hippopotamus (19), leopard (29), kangaroo (33), and raccoon (36) among BD than NC group participants. The opposite pattern was found for associations of elephant (5) with rabbit (18) and lizard (35).

A clearer discrepancy between NC and BD groups can be seen when whale (27) is compared with other animals, especially in higher dimensional spaces (dark and gray solid lines in Figure 2E). In healthy adults, whale (27) tends to show high association with snake (11), fish (16), mouse (17), rat (21), alligator (22), eagle (30), crocodile (31), shark (34), lizard (35), dolphin (28) and duck (39) when dimension 1–4 are considered (dark solid line). When Dimensions 1–5 are included (gray solid line), some of these, i.e. snake (11), mouse (17), rat (21), alligator (22), and crocodile (31) drop out, making the associations between whale (5) and the remaining animals (i.e., sea-animals cluster) more obvious. For BD, these subclusters do not seem to exist separately, even when the fifth dimension is considered.

In summary, animal clusters found in healthy adults could not be clearly identified in BD. The discrepancy was most obvious for wild/African animals and lower-ranked animal clusters (sea animals, reptiles), in which members are widely separated from each other and even overlap with other cluster exemplars.

Supermarket category: 2-D cluster representations.

The top 40 supermarket items named by NC and BD group members are plotted in Figure 3. As before, lower ranked exemplars (21–40) are shown in separate panels (Figure 3C and 3D). In Figure 3A and 3C for NCs, examples of fruits and vegetables tend to form a cluster on the right side of the 2-D plots (first and fourth quadrants). A cluster of staple/dairy products (milk, bread, ice cream, etc.) is mixed with subcategory names such as fruit, vegetable, and meat. The clusters described here can be more clearly identified when three or more dimensional vector spaces are examined. For example, the group of subcategory names (fruit, vegetable, candy, and meat) becomes more distinct from the dairy products cluster in higher dimensional space (see Supplementary Information). Also, a cluster of meat products (chicken, turkey, steak, and beef) that cannot be seen in Figure 3A and 3C emerges in a higher dimensional plot.

Figure 3.

Clustering of top 40 supermarket items in 2-D vector space (dimensions 2 and 3) are presented in four panels. (A) Supermarket items with ranks of 1 through 20 named by NCs (1–10 in black and 11–20 in gray; see Appendix for the full list of ranks). (B) Supermarket items (ranks 1–20) named by patients with BD. (C) Supermarket items with ranks of 21–40 named by NCs (21–30 in black and 31–40 in gray). (D) Supermarket items (ranks 21–40) named by patients with BD.

The clustering of these same supermarket items by patients with BD seems less coherent than that shown by NCs (Figure 3B and 3D). Specific vegetables and fruits are separated more widely in BD, and this incoherence seems to worsen when three or more dimensional spaces are considered (see next section). Some dairy products (eggs, cheese, and yogurt) are not clearly grouped with other products (ice cream, bread, cookies, etc.). Examples of meat products (chicken, turkey, etc.) are also mixed with dairy products even in three or more dimensional spaces. The cluster of subcategory names (fruit, meat, vegetables, and candy) seems to be visibly separated from other clusters, as was the case for NCs.

In short, healthy adults show very distinct clusters of supermarket items that resemble the way that items are organized and displayed in an actual supermarket. Higher dimensionality is needed to see the emergence of clear clusters. Patients with BD seem to show less coherent clustering of items. The vector angles between semantically related items are much wider in BD than in healthy adults and some clusters do not seem to exist in BD, even in higher-dimensional spaces. It is notable that the incoherency of semantic clusters shown by persons with BD appears unrelated to the frequency of words. In other words, persons with BD showed different clustering patterns for supermarket items regardless of the frequency ranks of words.

Supermarket items: cosines of vector angles.

The cosines of angles between three supermarket items: eggs (4), lettuce (8), pears (30), and the top 40 items are plotted in Figure 4. As before, the numbers in parentheses attached to each item indicate the frequency rank of the item. In Figure 4A for NC, eggs (4) tend to cluster clearly with other dairy products and related items: milk (1), bread (2), cheese (3), cereal (9), ice cream (10), butter (16), and so forth, as we have seen in Figure 3A and 3C. Some exceptions also were found, for example, chicken (7), beef (31), and yogurt (27), whose association patterns with eggs (4) are counterintuitive. Chicken (7) and beef (31), however, show stronger associations with other meat than with dairy products: fish (17), steak (25), ham (35), or turkey (38); see Supplementary Information.

Figure 4.

Cosines of selected word vectors against top 40 supermarket items (A) and (B). Cosines of angles between eggs (4) and the top 40 items by NC and BD, respectively (C) and (D). Cosines of vector angles between lettuce (8) and top 40 items by NC and BD. (E) and (F) cosines of vector angles between pears (30) and top 40 items by NC and BD. The horizontal axis represents top 40 supermarket items. They are, numbering by 5, milk (1), bread, cheese, eggs, apples (5), meat, chicken, lettuce, cereal, ice cream (10), oranges, soda, tomatoes, vegetables, potatoes (15), butter, fish, candy, bananas, fruit (20), carrots, cookies, cake, onions, steak (25), sugar, yogurt, soup, juice, pears (30), beef, toilet paper, bacon, potato chips, ham (35), coffee, celery, turkey, paper towels, lunch meat (40).

Patients with BD showed a somewhat different pattern of cosine values for eggs (4) across the top 40 items (Figure 4B). Some items that were strongly associated with eggs (4) in BD but not in NC are lettuce (8) and tomatoes (13). Ice cream (10), cookies (22), and cake (23) were very weakly associated with eggs (4) in persons with BD but strongly associated in healthy adults.

The cosines between lettuce (8) and the top 40 items by healthy adults (Figure 4C) are quite impressive because they do not fluctuate at all across three different dimensional plots (dotted, solid-dark, and solid-gray lines), indicating that these concepts are strongly and robustly associated. As expected, lettuce (8) is strongly associated with such fruits and vegetables as apples (5), oranges (11), tomatoes (13), potatoes (15), and so forth. One exception is onions (24), but this is probably because lettuce (8) and onions (24) are located at opposite boundaries of the same cluster in multidimensional space (see Supplementary Information). Note that subcategory names, e.g., vegetables (14) and fruit (20), are not the members of this cluster.

As already seen in Figure 3B and 3D, patients with BD showed very disturbed semantic associations for lettuce (8) and other vegetables and fruits (Figure 4D). Lettuce (8) showed strong associations with cheese (3), eggs (4), onions (24), yogurt (27), and beef (31) in BD. This is not seen in NCs. Nor did lettuce (8) show any associations with apple (5), oranges (11), and bananas (19).

Healthy adults show apprehensible clustering pattern between pears (30) and other vegetables and fruits (Figure 4E). The cosine values of pears (3) in Figure 4E are almost identical to those of lettuce (8) in Figure 4C. Unlike in Figure 4C, fish (17) and steak (25) do not show strong associations with pears (30) but onions (25) do, which makes clear semantic groups. In BD, it is quite difficult to grasp a meaningful association pattern between pears (30) and the top 40 items (Figure 4F). The pattern of cosine values does not give a clear picture of semantic associations among supermarket items in the BD group.

To summarize, consistent with the findings shown in Figure 3B and 3D, the pattern of cosines of angles between exemplar vectors confirm that patients with BD show much less coherent grouping of supermarket items than do healthy adults.

Discussion

Previous reports of decreased verbal fluency in persons with BD (Bearden et al., 2001; Dickerson et al., 2004; Quraishi & Frangou, 2002; Schretlen et al., 2007; Seidman et al., 2002) could not identify possible sources of the impairment, due to their reliance on measures of productivity. Consequently, the similarity between patients with BD and SZ reported by these studies has been superficial. Here we showed that the clustering patterns of retrieved concepts, which relates directly to the ways in which semantic concepts are stored and retrieved, differed between BD and NC groups despite comparable productivity. These findings shed light on the nature of semantic system impairment in persons with bipolar disorder.

Some have argued that decreased productivity and aberrant clustering of verbal fluency productions can result from the literal deterioration of semantic information itself (i.e., the storage deficit hypothesis) or any related structural component of the semantic system (Bozikas, Kosmidis, & Karavatos, 2005; Laws, Al-Uzri, & Mortimer, 2000; Paulsen et al., 1996; Rossell & David, 2006). Alternatively, the functional deficit hypothesis posits that decreased productivity and aberrant clustering can both result from a functional impairment of the activation/inhibition and related processes required to access or retrieve semantic information (Allen, Liddle, & Frith, 1993; Aloia et al., 1998; Joyce et al., 1996; Spitzer, Maier, & Weisbrod, 1997). These two explanations are not mutually exclusive, but the results of our SVD analysis are more consistent with a functional than structural explanation of semantic impairment in BD. We interpret the less coherent clustering of category exemplars by adults with BD as evidence of impaired semantic activation/inhibition, activation spreading, or possibly the control of these functions. Several findings of the present study support this interpretation.

First, a storage deficit, such as that seen in Alzheimer’s disease, is closely bound to reduced productivity resulting from the loss of semantic information. Further, because successfully named exemplars should be concepts that survive such degradation, a storage deficit alone does not lead to clear predictions about semantic clustering (e.g., Rossell & David, 2006). In our study, patients with BD showed intact productivity on animal naming, yet the semantic clustering of words they produced was fairly disturbed. The members of two major clusters (wild/African animals and farm/domestic animals) were more widely separated in the vector space representations of patients with BD than of NCs (see Figure 1). Also, unlike the NC group, patients in the BD group did not produce separate clusters of low-ranked animals (sea animals and reptiles). It is difficult to explain this incoherent clustering in the context of intact productivity by patients if a storage deficit is solely responsible for the semantic impairment in BD. Similar logic applies to the clustering of supermarket items. Patients with BD showed just slightly reduced output; the supermarket items they named formed fairly disturbed clusters, even for high-frequency words. Given the modest diminution in productivity, it is difficult to attribute the aberrant clustering of supermarket items to a storage deficit. It seems that at least an additional problem in the semantic system is required to explain these findings.

Second, we found that the four top-ranked animals (cat, dog, lion, and tiger) were generally clustered as domestic/farm animals (dog and cat) and wild/African animals (lion and tiger) by patients with BD but not by healthy adults. While this difference is obscure in 2-D plots for dimensions 2 and 3 (see Figure 1), it becomes clear when higher dimensional space is considered (see Figure 2). This interesting pattern has been observed in some other analyses, including our previous study of SZ (Aloia et al., 1996; Sumiyoshi et al., 2005; Sung et al., 2012). We propose that this counterintuitive phenomenon results from the exceedingly strong associations between these top-ranked exemplars and the category cue (“animals”), obviating the need for contextual guidance to retrieve them. Putting it differently, these four animal concepts simply pop out in response to the cue “animals.” When these strong associations or activation processes are compromised for some reasons, the retrieval of these concepts will require some contextual cueing, with the result that patients with both BD and SZ show more obvious clustering of these four concepts than healthy adults.

The disassociation between overall productivity and the pattern of semantic clusters we report here suggests that the claimed semantic deficit behind the altered word clusters in BD may be independent of what reduced productivity and related determinants, such as processing speed, tell us. That is, impaired semantic clusters may reveal an aspect of cognitive impairment in BD that is independent of reduced productivity and related deficits (e.g., Dickinson, Ragland, Calkins, Gold, & Gur, 2006). Along this line, we (Vannorsdall, Maroof, Gordon, & Schretlen, 2012) previously found that generation of words and ideas seems to form an unique cognitive dimension that is largely independent of processing speed and other cognitive abilities.

The cognitive processes that support semantic fluency probably cannot be fully captured by the single measure of overall productivity (Troyer et al., 1997). Nonetheless, given the proposed explanation for altered word clusters in BD, we expect that semantic priming (e.g., McNamara, 2005) would be impaired to a certain degree, and this has been found by a very limited number of studies (Andreou, Bozikas, Ramnalis, Giannakou, & Fokas, 2013; Ryu et al., 2012). At this point, the direct correlational relationship between impaired clustering and semantic priming cannot be examined due to the qualitative nature of the clustering results and the early stage of research on semantic priming in BD.

Another interesting observation is that, compared with our previous study, the clustering patterns shown by persons with BD appear to fall between those shown by healthy adults and persons with SZ. In this sense, the present findings provide further support for the inference that, in terms of neurocognitive functioning, the differences between SZ and BD are more quantitative than qualitative (Schretlen et al., 2007), which has a practical significance in defining and understanding of the disease.

Some limitations of the current study merit comments. First, participants did not provide mood ratings at the time of testing, so we cannot assess any effect of affective status on their task performance. However, the patients were relatively stable at the time of testing. Note that Rossell et al. (2006) found that patients with BD showed reduced productivity only in response to cues with affective valence, such as naming “happy” events or objects, and not in response to affectively neutral cues, such as naming animals or food items. Also, Dixon, Kravariti, Frith, Murray, and McGuire (2004) suggested that cognitive impairment in BD, which seemed to be differentially associated with different mood status (e.g., manic and depressed), may be generally related to the underlying cause of the disorder than a particular symptom profile. In other words, the fundamental cause for cognitive impairment in bipolar patients remains the same, regardless of the mood status of the patients and their performance variation. Similarly, Bora, Yucel, and Pantelis (2010) argued that the pattern of cognitive impairments shown by patients with affective psychoses was very similar to those shown by euthymic BD patients, just more severe. These considerations seem to decrease the likelihood that individual differences in mood at the time of testing made more than a modest contribution to the findings of this study.

Second, we did not characterize subtypes of BD beyond noting that most (88%) of the patients had BD I disorder. Studies show that patients with both BD I and II disorders can show cognitive deficits, although these generally are more severe in BD I than BD II (e.g., Chang et al., 2011; Hsiao et al., 2009; Simonsen et al., 2008; Torrent et al., 2006). Thus, the fact that we found significant differences in the clustering of category exemplars by a mixed sample of patients with BP I and II, works in favor of the conclusion that BD is characterized by a functional impairment of the semantic system. Similarly, we did not discriminate bipolar patients with and without psychosis in the analysis. This is mainly because patients with psychosis were not different from those without psychosis in terms of their productivity in two conditions. On average, the 57 patients with a history of psychotic symptoms named 19.2 animal names and the 34 patients without a history of psychosis¹ named 18.6 animals, t(89) = −0.42, p = .68. Patients with and without histories of psychosis named about the same number of supermarket items on average, 23.8 vs. 22.9, respectively; t(89) = −0.64, p = .54. Still, our sample size does not permit a direct comparison of subgroup clustering patterns. Finally, we have not provided any objective measure of goodness of fit for our vector solutions (e.g., 2-D representations). As stated in the introduction, our goal was to compare levels of clustering coherency of patients with BD and healthy controls, not to find best solutions for each group. We believe we achieved this goal without such measures. In any case, there is no generally accepted method to find the best multidimensional solution for SVD analysis (Quesada, 2007).

To conclude, it is fairly clear that reduced category-fluency output does not point to a specific source of semantic system dysfunction. Nor does semantic dysfunction inevitably reduce output productivity on verbal fluency tasks. In this study, SVD of category-fluency data revealed clearly disturbed clustering without impoverished output in adults with BD. Most important is that the source of this abnormality seems more closely linked to functional aspects of the semantic system than to any degradation of semantic knowledge.

Supplementary Material

http://dx.doi.org/10.1037/a0033068.supp

Table 2.

Descriptive Statistics of Category Fluency Performance

Group	Category	Mean (SD)	Min/Max
NC	Animals	20.6 (5.6)	7/36
	Supermarket	26.0 (7.2)	12/44
BD	Animals	19.1 (5.6)	9/36
	Supermarket	23.3 (6.5)	10/42

Appendix

Frequency Ranks of Animal Names and Supermarket Items. Words Are Sorted by the Frequencies Calculated from a Verbal Fluency Database of 780 Adults (Both Patients and Healthy Controls), Including Participants Who Contributed Data to the Current Analysis

Rank	Animals	All	BD	NC	Supermarket items	All	BD	NC
1	cat	718	82	91	milk	581	76	75
2	dog	713	81	92	bread	503	67	63
3	lion	610	69	85	cheese	426	51	65
4	tiger	550	59	71	eggs	368	45	54
5	elephant	519	54	70	apples	323	40	37
6	giraffe	406	56	52	meat	310	32	35
7	bear	403	43	60	chicken	297	33	35
8	horse	395	51	49	lettuce	291	37	41
9	zebra	379	51	52	cereal	288	38	40
10	monkey	346	53	50	ice cream	282	38	46
11	snake	342	54	43	oranges	275	34	35
12	cow	339	42	44	soda	254	35	37
13	bird	305	42	39	tomatoes	230	22	31
14	pig	235	28	34	vegetables	216	22	25
15	deer	207	19	29	potatoes	215	17	31
16	fish	206	29	29	butter	213	25	25
17	mouse	193	22	28	fish	211	26	19
18	rabbit	192	17	32	candy	208	27	28
19	hippopotamus	190	32	26	bananas	202	27	30
20	rhinoceros	176	22	16	fruit	191	24	22
21	rat	168	20	20	carrots	176	21	32
22	alligator	161	23	15	cookies	175	18	26
23	squirrel	159	12	22	cake	169	21	17
24	gorilla	153	25	22	onions	165	16	21
25	sheep	152	13	27	steak	162	20	17
26	chicken	152	15	27	sugar	157	15	13
27	whale	147	22	23	yogurt	144	24	22
28	goat	139	18	20	soup	138	20	14
29	leopard	138	13	10	juice	133	17	25
30	eagle	115	13	13	pears	127	17	12
31	crocodile	110	18	13	beef	126	19	17
32	fox	109	7	19	toilet paper	126	15	26
33	kangaroo	107	11	13	bacon	124	13	23
34	shark	104	24	12	potato chips	124	15	20
35	lizard	102	18	11	ham	123	17	15
36	raccoon	101	7	14	coffee	122	7	19
37	ape	93	14	10	celery	122	11	16
38	dolphin	88	13	13	turkey	121	16	15
39	duck	87	11	10	paper towels	119	7	23
40	donkey	85	16	11	lunch meat	110	16	22