Beyond Diagnosis: Transdiagnostic Cognitive Subgroups in Mood Disorders Linked to Work‐Related Disability

ABSTRACT

Aim

Cognitive dysfunction is a transdiagnostic feature of mood disorders and a key determinant of functional outcomes. This study aimed to identify transdiagnostic cognitive heterogeneity among patients with mood disorders using a data‐driven approach and to examine its association with social functioning. Characterizing cognitive subtypes and their association with subjective functional disability is an essential first step toward stratified interventions and personalized care.

Methods

We analyzed data from 264 age‐ and sex‐matched patients (132 each with major depressive disorder [MDD] or bipolar disorder [BD]). Participants completed six neurocognitive tests to assess executive function, memory, attention, and processing speed. Cognitive variables were adjusted for age and sex, and unsupervised K‐means clustering was used to identify transdiagnostic subgroups. Functional impairment was evaluated using the Sheehan Disability Scale.

Results

After matching participants based on age and sex, we compared the cognitive performance between the diagnostic groups. Significant differences were observed in only two cognitive variables with a small effect size (mean |Cohen's d| = 0.14). No differences were found in the social functioning measures. Cluster analysis identified two cognitive subtypes: higher‐performance group (n = 200) and lower‐performance group (n = 64) with a large effect size (mean |Cohen's d| = 0.99). Both clusters included patients with MDD and BD in similar proportions (p = 0.11), indicating a diagnostic overlap. The lower‐performance group also showed significantly greater work‐related and total functional impairments.

Conclusion

Unsupervised clustering revealed transdiagnostic cognitive subgroups in mood disorders independent of diagnostic labels. Cognitive phenotyping may inform personalized treatment strategies and improve functional outcomes by identifying high‐risk patients for cognitive remediation or neuromodulation therapies.

Diagnostic categories alone did not explain differences in subjective social functioning, whereas cognitive clustering identified a lower‐performance subgroup with worse work and total disability. Cognitive heterogeneity may underlie functional impairment across mood disorders.

1. Introduction

Major depressive disorder (MDD) and bipolar disorder (BD) are among the most prevalent disabling psychiatric disorders [1]. Beyond mood symptoms, cognitive dysfunction is a prominent feature of mood disorders and an important determinant of real‐world functioning [2]. Cognitive impairment may persist even after symptomatic improvement and has been associated with poorer occupational and psychosocial outcomes [2, 3]. Work‐related impairment is especially clinically relevant because it directly affects independence, social recovery, and quality of life [4, 5]. However, cognitive performance varies widely across individuals, and categorical diagnoses alone do not adequately capture this heterogeneity [3, 6, 7].

Data‐driven approaches such as clustering analysis offer an opportunity to redefine subtypes based on objective cognitive profiles rather than clinical diagnostic labels [8, 9]. Previous studies [10, 11, 12] have demonstrated the value of subgrouping approaches across mood and other psychiatric disorders, but important gaps remain. One study derived subgroups based on functioning and quality‐of‐life profiles, with cognition examined only as a correlate rather than as the basis for clustering [10]. Another study identified neuropsychological clusters across MDD and BD but primarily focused on cognitive profiling and demographic or clinical correlates, without clearly validating the resulting subgroups against an independent, clinically actionable functional outcome [11]. More recent transdiagnostic work has linked neurocognitive subgroups to functional course over time [12], but this study was conducted in broader and younger samples spanning multiple emerging mental disorders, limiting their direct applicability to clinically comparable mood disorder cohorts and to domain‐specific outcomes such as work‐related disability.

In the present study, we used unsupervised clustering of 16 neurocognitive measures to identify cognitive subgroups in an age‐ and sex‐matched cohort of patients with MDD and BD. We then examined whether these subgroups differed in subjective functional disability, particularly work‐related impairment, assessed independently from the clustering procedure. We hypothesized that transdiagnostic subgroups would emerge that differ in both cognitive performance and functional outcomes.

2. Methods

2.1. Participants and Data Collection

This study was approved by the Local Ethics Committee of Hokkaido University (approval number: 024–0326). Participants were recruited from among individuals who visited the Department of Psychiatry at Hokkaido University Hospital between 2005 and 2019. All included patients were diagnosed with MDD or BD by board‐certified psychiatrists according to the ICD‐10 criteria and had completed all six standardized neurocognitive tests (see Methods; 2.2 Neurocognitive Assessment) without any missing data. Consequently, 187 patients with MDD and 132 patients with BD were included in this study. To control for demographic differences, propensity score matching was performed based on age and sex, resulting in a final sample of 264 individuals (132 with MDD and 132 with BD).

2.2. Neurocognitive Assessment

Participants completed six standardized neurocognitive tests administered by trained psychologists during a clinically stable phase, as judged by treating psychiatrists. The details of the battery are described by Toyomaki et al. [13]. A summary of the battery is provided below. The 16 selected cognitive variables are indicated in parentheses, each accompanied by a number.

1. Wisconsin card sorting test (WCST): A computerized Japanese Keio University version was used [14]. This task assesses cognitive flexibility and set‐shifting. In the present study, categories achieved were interpreted as an index of successful rule acquisition, whereas perseverative errors reflected difficulty in shifting from a previously learned rule. Perseverative errors were further examined as Milner‐type and Nelson‐type indices: Milner‐type errors indicate persistent responding according to a previously correct sorting principle, while Nelson‐type errors more specifically capture failure to shift in response to feedback that the sorting principle has changed. The number of categories achieved (1. WCST_CA) and the perseverative errors of Nelson (2. WCST_PEN) and Milner (3. WCST_PEM) were used to measure executive function.

2. Continuous performance test (CPT): A computerized A–X CPT was administered for approximately 7 min. Several characters were presented at the center of a display. The participants were instructed to respond as quickly as possible to “X,” which appeared immediately after “A.” The target stimulus was presented 70 times. Each stimulus was presented for 100 ms with an inter‐stimulus interval of 1500–2000 ms. The average reaction time (4. CPT_RT) was used to measure psychomotor speed. The total number of errors was used as a measure of attention (5. CPT_Error).

3. Word fluency test (WFT): each participant had to speak as many words as possible beginning with the “Shi,” “I,” or “Re” sounds in 60 s. The total number of words was used to measure verbal fluency (6. WFT_total), which is a part of executive function.

4. Stroop test: This test contains 24 dots printed in different colors or 24 color words; each printed in a different incongruent color. Each participant was instructed to name the printed colors of the dots and words as quickly as possible. The number of errors (7. Stroop_error) and reaction time difference (8. Stroop_dif) between the naming colors of dots and words reflect response inhibition and selective attention.

5. Trail Making Test (TMT): The TMT consisted of two parts, Parts A and B, each comprising 25 circles distributed over a sheet of paper. In Part A, the circles were numbered 1–25. The participants were required to draw lines connecting the numbers in ascending order. In Part B, the circles included both numbers (1–13) and 12 hiragana (Japanese cursive syllabary) letters, and the participants were required to draw lines with the added task of alternating the numbers and letters. Part A measured the psychomotor speed and attention (9. TMT_A), whereas Part B required greater mental flexibility (10. TMT_B), the ability to shift attention, and strategy. Therefore, the subtracted time (11. TMT_BA: TMT‐Part B minus TMT‐Part A) was also used to measure executive function.

6. Auditory Verbal Learning Test (AVLT): Each participant was required to learn a 10‐item word list over four trials (trials 1–4); the participants were then assessed 30 min later (trial 5). The number of words recalled in trials 1–4 indicated immediate recall (12. VLT_1, 13. VLT_2, 14. VLT_3, 15. VLT_4). The number of words recalled in trial 5 indicated delayed recall (16. VLT_5).

2.3. Functional Assessment

Functional impairment was assessed using the Sheehan Disability Scale (SDS) [15], which is a brief self‐report instrument that measures disruptions in three domains: work/school, social life, and family/home responsibilities. Patients rated the extent to which their symptoms interfered with each domain using a 10‐point visual analog scale ranging from 0 (no impairment) to 10 (extreme impairment), with ratings of 1–3 indicating mild, 4–6 moderate, and 7–9 marked impairments. Each item was introduced with the question “Have your symptoms disrupted your [work/school, social life, or family life/home responsibilities]?” Total scores ranged from 0 to 30, with higher scores reflecting greater overall functional impairment.

2.4. Clustering and Data Analysis

All cognitive variables were first regressed on age and sex using linear regression, and standardized residuals were extracted to control for demographic effects. K‐means clustering was then applied to these residuals. Models with 2–6 clusters (K = 2–6) were evaluated, and the optimal solution was determined based on the silhouette score. We used the scikit‐learn implementation (version 1.6.1), specifying random_state = 42 and n_init = 10, while all other parameters were maintained at their default values. To assess solution stability with respect to random initialization, we repeated K‐means clustering 50 times using different random seeds (random_state = 100–149). Cluster labels were aligned to a reference solution using the Hungarian algorithm prior to computing agreement metrics. Stability was quantified using the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI), reported as mean ± SD. Group comparisons were conducted using independent‐samples t‐tests and chi‐squared tests, with a significance threshold of α = 0.05. The effect sizes (Cohen's d) were calculated. All statistical analyses were performed using Python (version 3.11.8).

3. Results

3.1. Diagnostic Comparison

First, we compared the MDD and BD groups after applying age‐ and sex‐based propensity score matching (Table 1). Significant group differences were observed in two cognitive variables: WCST_PEM and Stroop_dif. However, the effect sizes were generally small across most measures (average Cohen's d = 0.14). The results for the original (pre‐matched) sample are presented in Supplementary Table S1.

TABLE 1.

Demographic characteristics and neurocognitive properties (post‐matching).

	MDD (mean ± SD)	BD (mean ± SD)	t‐value	p‐value	|Cohen's d|
n (number)	n = 132	n = 132	−	−	−
age	42.17 ± 12.69	42.14 ± 12.76	0.02	0.98	0.00
sex	71 (53.8%)	70 (53.0%)	−	1.00	−
WCST_PEM	1.92 ± 3.28	2.98 ± 4.88	−2.09	0.04	0.26
WCST_PEN	3.78 ± 4.57	4.87 ± 5.80	−1.70	0.09	0.21
WCST_CA	4.69 ± 1.51	4.36 ± 1.74	1.63	0.10	0.20
CPT_Error	3.88 ± 4.92	4.47 ± 5.95	−0.88	0.38	0.11
CPT_RT	420.56 ± 100.45	432.02 ± 105.83	−0.90	0.37	0.11
WFT_total	27.17 ± 9.60	25.77 ± 8.80	1.24	0.22	0.15
Stroop_dif	14.72 ± 3.53	16.14 ± 6.97	−2.09	0.04	0.26
Stroop_error	0.67 ± 1.80	0.77 ± 1.31	−0.51	0.61	0.06
TMT_A	88.64 ± 28.84	92.47 ± 33.34	−1.00	0.32	0.12
TMT_B	109.17 ± 56.68	115.60 ± 54.52	−0.94	0.35	0.12
TMT_BA	20.53 ± 39.04	23.13 ± 40.68	−0.53	0.60	0.07
VLT_1	4.88 ± 1.52	4.83 ± 1.51	0.28	0.78	0.03
VLT_2	6.75 ± 1.70	6.67 ± 1.77	0.39	0.70	0.05
VLT_3	7.83 ± 1.58	7.46 ± 1.90	1.70	0.09	0.21
VLT_4	8.13 ± 1.56	7.89 ± 1.71	1.21	0.23	0.15
VLT_5	6.17 ± 2.45	5.82 ± 2.62	1.12	0.27	0.14
SDS_work	5.38 ± 3.11	5.42 ± 3.14	−0.12	0.91	0.01
SDS_social	4.77 ± 3.07	4.91 ± 3.11	−0.38	0.71	0.05
SDS_fam	3.61 ± 2.99	4.22 ± 3.20	−1.61	0.11	0.20
SDS_total	13.75 ± 7.98	14.55 ± 8.39	−0.80	0.43	0.10

3.2. Cluster Analysis

Data‐driven approaches enable the identification of novel subtypes based on objective cognitive profiles. We proceeded with a transdiagnostic clustering approach using all age‐ and sex‐regressed 16 cognitive variables (see Methods). Silhouette analysis indicated that a two‐cluster solution (K = 2) provided the best fit (Figure 1A), yielding two distinct groups: Class 1 (n = 200) and Class 2 (n = 64) (Figure 1C). To aid in interpretation, we referred to Class 1 as the higher‐performance group and Class 2 as the lower‐performance group. These labels indicated relative performance within the patient sample and do not imply intact cognition relative to normative data or healthy controls. The higher‐performance group included 106 patients with MDD (53%) and 94 with BD (47%), whereas the lower‐performance group included 26 patients with MDD (40.6%) and 38 with BD (59.4%) (Figure 1B). There was no significant difference in diagnostic composition between clusters (χ² = 2.50, p = 0.11). Across all 16 cognitive variables, the lower‐performance group performed significantly worse than the higher‐performance group (Table 2), with large effect sizes across most measures (average Cohen's d = 0.99). Furthermore, the lower‐performance group exhibited significantly greater functional impairment on the SDS work and total subscale scores (Table 2). The two‐cluster solution showed high stability across repeated random initializations (ARI = 0.999 ± 0.004; NMI = 0.998 ± 0.009).

FIGURE 1.

Results of cluster analysis. (A) Silhouette scores for solutions with 2 to 6 clusters. The highest score was observed for a two‐cluster solution. (B) Diagnostic composition of each cluster (higher‐performance group vs. lower‐performance group). Red and blue bars represent the proportion of patients with MDD and BD, respectively. (C) Visualization of the two clusters in a two‐dimensional PCA space. Green circles/crosses and orange circles/crosses denote higher‐performance group and lower‐performance group, respectively. Green/orange circles indicate MDD patients; green/orange crosses indicate BD patients. Ellipses show 95% confidence regions based on the covariance of PCA scores within each cluster.

TABLE 2.

Comparison of neurocognitive properties in clustered populations.

	Higher (mean ± SD)	Lower (mean ± SD)	t‐value	p‐value	|Cohen's d|
n (number)	n = 200	n = 64	−	−	−
age	41.62 ± 11.92	43.84 ± 14.86	−1.09	0.28	0.17
sex	103 (51.5%)	38 (59.4%)	−	0.34	0.06
WCST_PEM	1.42 ± 2.31	5.69 ± 6.50	−5.16	2.32 × 10−6	0.88
WCST_PEN	2.98 ± 3.51	8.53 ± 7.20	−5.94	8.91 × 10−8	0.98
WCST_CA	4.96 ± 1.29	3.16 ± 1.84	7.33	1.35 × 10−5	1.14
CPT_Error	3.25 ± 4.40	7.08 ± 7.21	−4.02	1.32 × 10−4	0.64
CPT_RT	406.45 ± 87.45	488.30 ± 122.86	−4.94	3.85 × 10−6	0.77
WFT_total	28.20 ± 9.03	21.06 ± 7.64	6.21	7.15 × 10−9	0.85
Stroop_dif	14.44 ± 5.10	18.53 ± 5.83	−5.03	2.25 × 10−6	0.75
Stroop_error	0.48 ± 0.94	1.45 ± 2.61	−2.93	4.64 × 10−3	0.50
TMT_A	83.26 ± 24.66	113.36 ± 37.92	−5.96	6.34 × 10−8	0.94
TMT_B	96.56 ± 37.12	161.86 ± 72.59	−6.91	1.44 × 10−9	1.13
TMT_BA	13.29 ± 28.25	48.50 ± 56.07	−4.83	7.23 × 10−6	0.79
VLT_1	5.19 ± 1.40	3.80 ± 1.36	7.08	1.47 × 10−10	1.01
VLT_2	7.15 ± 1.52	5.33 ± 1.64	7.86	4.59 × 10−12	1.15
VLT_3	8.19 ± 1.32	5.94 ± 1.83	9.10	3.40 × 10−14	1.41
VLT_4	8.53 ± 1.19	6.38 ± 1.76	9.15	3.50 × 10−14	1.43
VLT_5	6.75 ± 2.08	3.61 ± 2.37	9.50	2.00 × 10−15	1.41
SDS_work	5.12 ± 3.10	6.30 ± 3.02	−2.71	7.92 × 10−3	0.39
SDS_social	4.66 ± 3.06	5.39 ± 3.12	−1.64	0.10	0.24
SDS_fam	3.75 ± 3.07	4.41 ± 3.19	−1.43	0.15	0.21
SDS_total	13.53 ± 8.15	16.09 ± 8.05	−2.21	0.03	0.32

4. Discussion

In this study, we applied an unsupervised clustering approach to a transdiagnostic sample of patients with MDD and BD, using performance on a comprehensive battery of neurocognitive tests. Our analysis identified two distinct subgroups: a higher‐performance group, characterized by relatively preserved cognitive function, and a lower‐performance group, exhibiting marked impairments across all assessed domains. Importantly, these clusters were not defined by diagnostic boundaries; both patients with MDD and BD were represented in each group, and the diagnostic composition did not significantly differ between the clusters (Figure 1B). This finding supports the growing evidence that mood disorders share overlapping neurocognitive variables and that cognitive profiles may cut across traditional diagnostic categories¹¹.

Although the two‐cluster solution was derived from differences in the 16 neurocognitive variables, between‐group cognitive differences were expected by design. The more clinically important finding was that the lower‐performance group also showed greater functional disability, particularly in work‐related domains, even though functional measures were not included in the clustering procedure. This supports the external clinical relevance of the identified cognitive subgroups and is consistent with previous evidence linking cognitive impairment to real‐world functional outcomes [16].

Taken together, these findings extend the existing literature in four important ways. First, we identified transdiagnostic cognitive subgroups within a clinically relevant MDD–BD cohort in which age and sex were carefully matched, thereby minimizing demographic confounding between diagnostic categories. Second, we used a standardized multidomain cognitive battery (16 variables) to capture cognitive heterogeneity within mood disorders. Third, we provided external clinical validation by showing that the identified cognitive subgroups differed in self‐reported functional disability, particularly work‐related impairment, which was assessed independently of the clustering procedure. Fourth, we enhanced methodological rigor by reporting both cluster selection and stability (e.g., silhouette‐based model selection and stability metrics across repeated initializations), thereby addressing concerns regarding arbitrariness and reproducibility in clustering‐based research.

Cognitive subtyping may inform individualized treatment approaches. Patients in the cognitively impaired subgroup may benefit from targeted cognitive remediation [17], neuromodulation strategies [16], or biofeedback‐based interventions [18]. Early identification of such subtypes could aid in the prognosis and optimization of resource allocation in clinical settings.

The present study has several limitations. First, several potentially important clinical and background variables, including symptom severity, illness duration, medication status, educational attainment, and premorbid intellectual ability, were not systematically available in the present dataset. This limited our ability to further characterize the identified cognitive subgroups and may have contributed to between‐subgroup differences. These factors may also have influenced the subgroup structure itself. Therefore, although the two‐cluster solution showed the best fit and stability in the present dataset, more diverse subtypes may emerge in larger and more comprehensively characterized samples. Previous studies have reported three to four clusters, which may be partly attributable to their inclusion of a wider range of clinical and background variables than were available in the present study [10, 11, 12]. In future studies, adjusting for overall cognitive performance may help identify subtypes characterized more specifically by domain‐level differences. In addition, incorporating a broader range of clinical and biological parameters alongside cognitive variables may enable the identification of more diverse subtypes. Second, an additional limitation is the absence of a healthy control group. Although age and sex were adjusted within the patient sample to reduce demographic confounding, this approach does not provide a normative benchmark for interpreting the absolute severity of cognitive impairment. Accordingly, the identified subgroups should be understood as representing relative differences within the MDD–BD cohort rather than preserved versus impaired cognition in relation to healthy individuals. Third, information on BD subtype (type I vs. II) and biological measures was unavailable, which may have obscured diagnosis‐specific effects and limited mechanistic interpretation. Prior work suggests that cognitive profiles may vary across bipolar subtypes and may also be associated with biological markers [19, 20].

5. Conclusion

Using a data‐driven clustering approach, we identified two cognitive subtypes in mixed populations of MDD and BD. The presence of a transdiagnostic subgroup of patients with severe cognitive and functional impairments underscores the value of cognitive phenotyping in the field of psychiatry. Future studies should investigate the neurobiological underpinnings and longitudinal trajectories of these cognitive subtypes.

Author Contributions

R.S. contributed to conceptualization, methodology, formal analysis, visualization, project administration, original draft preparation, and manuscript revision. T.T. contributed to formal analysis and supervision of the analytical process. H.W. contributed to supervision and manuscript revision. A.T. contributed to conceptualization, supervision and manuscript revision. R.O. contributed to supervision. N.H. contributed to conceptualization, supervision and manuscript revision. T.A.K. contributed to supervision. All authors reviewed and approved the final manuscript.

Funding

This work was supported by Japan Society for the Promotion of Science, 24K10699.

Ethics Statement

This study was approved by the Local Ethics Committee of Hokkaido University (approval number: 024–0326) and was conducted in accordance with the principles of the Declaration of Helsinki and its later amendments.

Consent

In accordance with institutional guidelines for retrospective studies, written informed consent was waived and information about the study was disclosed publicly to allow patients the opportunity to opt out.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Demographic characteristics and neurocognitive properties (pre‐matching).

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.