The centrality of working memory networks in differentiating bipolar type I depression from unipolar depression: A task-fMRI study

Abstract

Objectives

Up to 70%–80% of patients with bipolar disorder are misdiagnosed as having major depressive disorder (MDD), leading to both delayed intervention and worsening disability. Differences in the cognitive neurophysiology may serve to distinguish between the depressive phase of type 1 bipolar disorder (BDD-I) from MDD, though this remains to be demonstrated. To this end, we investigate the discriminatory signal in the topological organization of the functional connectome during a working memory (WM) task in BDD-I and MDD, as a candidate identification approach.

Methods

We calculated and compared the degree centrality (DC) at the whole-brain voxel-wise level in 31 patients with BDD-I, 35 patients with MDD, and 80 healthy controls (HCs) during an n-back task. We further extracted the distinct DC patterns in the two patient groups under different WM loads and used machine learning approaches to determine the distinguishing ability of the DC map.

Results

Patients with BDD-I had lower accuracy and longer reaction time (RT) than HCs at high WM loads. BDD-I is characterized by decreased DC in the default mode network (DMN) and the sensorimotor network (SMN) when facing high WM load. In contrast, MDD is characterized by increased DC in the DMN during high WM load. Higher WM load resulted in better classification performance, with the distinct aberrant DC maps under 2-back load discriminating the two disorders with 90.91% accuracy.

Conclusions

The distributed brain connectivity during high WM load provides novel insights into the neurophysiological mechanisms underlying cognitive impairment of depression. This could potentially distinguish BDD-I from MDD if replicated in future large-scale evaluations of first-episode depression with longitudinal confirmation of diagnostic transition.

Introduction

Distinguishing bipolar disorder from major depressive disorder (MDD) in individuals is one of the most clinically important challenges in psychiatry. ¹ Accurate early diagnosis of bipolar disorder is difficult in the clinic, as a depressive episode is usually its first manifestation,^2,3 and subthreshold manic symptoms are often missed in the clinic. ⁴ Misdiagnosis of depression in bipolar disorder(BDD) as MDD is reported in up to 70%–80% of cases. ⁵ This leads to inappropriate medication, accelerated switch to mania, elevated suicidality, and poor prognosis especially in the bipolar type I depression (BDD-I).^1,6,7 Prior studies have shown that patients with established bipolar disorder exhibit more cognitive impairments than patients with MDD, and neurocognitive dysfunction independently contributes to poor functional outcomes in patients. ⁶ Thus, physiological markers related to cognitive function may be differentially altered in these two disorders, and may assist in differentiating BDD-I from MDD much earlier in the disease course.

Of the various neurocognitive deficits, working memory (WM) deficit, which describes the failure to represent, maintain, and update information in a short period, is a feature persistently seen in patients with BDD and MDD.^8,9 Previous studies have pointed out that patients with bipolar disorder and MDD demonstrated significant deficits in almost all domains of cognitive function including WM, whereas patients with bipolar disorder underperform patients with MDD only in WM and cognitive inhibition tests.^8,9 Other studies also reported greater impairment in patients with bipolar disorder in measurements of verbal memory, attention, and executive function relative to patients with MDD.^10,11 Hence, differences in the neural network underlying WM can be expected in patients with BDD and MDD.

In healthy controls (HCs), a distributed set of regions participate in WM functions during tasks like the n-back task; these areas include the medial prefrontal cortex (mPFC), the dorsolateral prefrontal cortex (DLPFC), the anterior cingulate cortex (ACC), the posterior parietal cortex, and medial temporal regions.^12,13 These regions constitute distinct large-scale brain networks with a pattern of WM-related activation seen in the central executive network (CEN), the salience network (SN), and deactivation in the default mode network (DMN). ¹² A failure to deactivate the DMN during WM task has been reported both in bipolar disorder and MDD participants.^14,15 Furthermore, a decrease in functional connectivity (FC) in the frontoparietal network during WM demand has been reported only in patients with MDD. ¹⁶

A small number of brain imaging studies have directly compared patients with bipolar disorder and MDD during the performance of cognitive tasks. Rodríguez-Cano et al. compared 26 BDD-I patients, 26 MDD patients, and 26 HCs in a 2-back versus baseline contrast, reporting a failure to deactivate the DMN in patients with BDD-I and MDD relative to the controls. This DMN dysfunction was more severe in BDD-I than MDD. ¹⁷ Rive et al. compared 19 MDD patients and 9 remitted bipolar disorder patients during performance of the visuospatial planning task. An increased frontostriatal activation during planning was reported as a specific characteristic of bipolar disorder. ¹⁸ However, despite these promising reports, the clinical utility of these WM-related neurophysiological differences in discriminating between these two mood disorders remains unknown, partly due to the lack of an easily accessible summary metric to quantify a spatially distributed pattern of network-level changes.

Recently, degree centrality (DC), a graph theory-derived measurement of network organization reflecting the number of instantaneous functional connections between a region (voxel) with the rest of the brain regions (voxel), has emerged as a measure that offers new insights into the patterns of FC in various diseases.^19,20 We have previously reported that in bipolar disorder, a shift in the distribution of DC occurs within the functional connectome during a WM task. ²¹ The distributed pattern of DC from resting functional magnetic resonance imaging (fMRI) has been already shown to carry diagnostic information, separating bipolar from unipolar depression with a high degree of accuracy (86%). ²² However, to the best of our knowledge, no study has estimated the utility of the DC pattern during cognitive tasks when differentiating BDD-I from MDD.

In this study, we acquired brain imaging data from patients with BDD-I, MDD, and HCs while they performed the n-back task. Then, we calculated and compared the whole-brain voxel-wise DC pattern of functional networks in all groups across different WM loads, and used these patterns in the subsequent classification analyses. Our objective is to demonstrate that when compared to healthy subjects, patients in a depressed state with an established diagnosis of BDD-I exhibit different patterns of DC abnormalities than those with a diagnosis of MDD, when performing WM task, especially at higher loads of cognitive demand. This objective, if met, can guide future studies that aim to estimate the risk of developing bipolar disorder at the first episode of depression.

Materials and Methods

Participants

A total of 161 subjects participated in this study, including 35 patients with BDD-I, 38 patients with MDD, and 88 HCs. The patients were all recruited from the Second Xiangya Hospital of Central South University and interviewed by two experienced psychiatrists. All patients were experiencing an episode of depression. To ensure the stability of the diagnosis of MDD subjects, so that BDD-I are not included in the MDD group, we traced the clinical profile of MDD patients from the time of the MRI data acquisition for a period of 2 years in an ambulatory setting (using case notes of treating clinicians), and ensured that no diagnostic uncertainty (i.e., conversion from MDD to BDD) occurred in this timeframe. The HCs were recruited from the community and assessed with the SCID-I/NP, ensuring that both HCs and their first-degree relatives had no current medical problems or family history of psychiatric disorders. Details of inclusion criteria and exclusion criteria for the participants are shown in Supplemental Material S1.

The clinical assessments of the patients with BDD-I and MDD included the Hamilton Depression Rating Scale (HAMD) and the Young Mania Rating Scale (YMRS).^23,24 Cognitive function was evaluated by the Wechsler Adult Intelligence Scale (WAIS). ²⁵

This study was approved by the Ethics Committee of Second Xiangya Hospital, Central South University. All participants provided written informed consent at the time of enrollment.

Behavioral Paradigm

As in our prior works,^26,27 we adopted the WM paradigm that comprised “0-back” and “2-back” load conditions in this study. A detailed description of this paradigm is given in Supplemental Material S2 and Figure S1.

MRI Data Acquisition and Preprocessing

Details of the neuroimaging data acquisition and preprocessing are described in the Supplemental Materials S3–S5.^28,29 After quality control procedures, a total of 31 patients with BDD, 35 patients with MDD, and 80 HCs were included in the final analysis. No significant difference was found in framewise displacement across the three groups (F_2,145 = 0.87, P = 0.42).

DC Calculation

Voxel-wise DC was computed using DPARSF toolbox (http://rfmri.org/DPARSF). The time course of each voxel was extracted and correlated with every other voxel in the brain to generate a correlation matrix, and the correlation matrix was binarized by thresholding at r = 0.25 before counting the number of connections to generate voxel-wise DC. For each participant, a DC map with DC values for every voxel was obtained, and these maps were normalized to z maps using the mean value and standard deviation within the whole gray matter mask. ³⁰

For each subject, the DC map was computed as the number of connections for each voxel, while, for some participants, there was no voxel that survived after the correlation matrix was binarized by thresholding at r > 0.25. Hence, these participants were excluded from group comparisons. The remaining data for further analysis during each condition came from: 25 patients with BDD, 29 patients with MDD, and 59 HCs during resting periods; 28 patients with BDD, 33 patients with MDD, and 79 HCs during 0-back WM load; 26 patients with BDD, 29 patients with MDD, and 80 HCs during 2-back WM load.

To validate whether the main results depended on the selection of correlation thresholds, we also calculated the DC maps using other different correlation thresholds (i.e., 0.15 and 0.35; see Supplemental Material S6).

Statistical Analysis

Statistical analysis of demographic and clinical data was performed using SPSS 18.0. To compare the differences among the three groups, one-way analysis of variance (ANOVA) or t-test was used for continuous variables and a chi-square test for categorical variables. Reaction time (RT) and accuracy were analyzed by three (group: BDD vs. MDD vs. HCs) × 2 (task: 0-back load vs. 2-back load) repeated measures ANOVA with group as a between-subject factor and task as a within-subject factor.

The statistical analysis of fMRI data was performed using Statistic Parameter Mapping 12 software (www.fil.ion.ucl.ac.uk/spm). First, one-way ANOVA was used to compare the DC maps among the three groups with gender, age, and education years as covariates, and a significant difference was set as a significance threshold of P < 0.005 at the cluster level. Then, by applying the significant voxels in ANOVA results as the mask, post hoc t-tests were performed between each two-group pair using a threshold at P < 0.05 with false discovery rate (FDR) correction to explore the difference in voxel-wise network centrality, in which gender, age, and education years were included as covariates as well.

Exploratory Analysis

Correlation analysis: Correlation analysis was employed to test the relationship between the mean values of altered DC in regions detected to have discriminatory information between BDD-I and MDD with the defined daily doses of medications, clinical symptoms (the score of HAMD and YMRS), cognitive function (the score of WAIS) and WM performance. The multiple comparisons in the correlation analysis were corrected with FDR (adjusted P < 0.05). A detailed description of the correlation analysis is given in Supplemental Material S7.

Machine learning analysis: We extracted the mean DC of regions that showed significant group differences between BDD-I and MDD groups during resting periods, 0-back load, and 2-back WM load, respectively, and conducted the subsequent classification using the support vector machine (SVM) toolkit LIBSVM (http://www.csie.ntu.edu.tw/∼cjlin/libsvm/). A detailed description of the SVM classification is given in Supplemental Material S8.

Results

Participant Characteristics

Detailed demographic, clinical characteristics, cognitive function, and behavioral data of participants are shown in Table 1. One-way ANOVA revealed a significant difference in age (F_2,145 = 27.67, P < 0.001), education (F_2,145 = 6.69, P = 0.002), scores on the Information subscale of WAIS (F_2,145 = 4.73, P = 0.01), and scores on the Digit subscale of WAIS (F_2,145 = 40.16, P < 0.001) across the three groups. Age, gender, and education were added as covariates in further analysis to address confounding effects.

Table 1.

Demographic and Clinical Characteristics, Cognitive Function, and Behavioral Data among Three Groups.

	BDD-I (n = 31)	MDD (n = 35)	HCs (n = 80)	F/T/χ2	df	P values	Post Hoc
Sociodemographic variable and clinical characteristics
Age (years)	24.74 (5.33)	33.03 (9.63)	23.46 (4.93)	27.67	2	< 0.001	MDD>BDD-I (P < 0.001); MDD > HCs (P< 0.001)
Sex (Male/Female)	12/19	20/15	38/42	2.25	2	0.32	N/A
Education (years)	13.48 (2.91)	12.10 (2.69)	14.04 (2.46)	6.69	2	0.002	BDD-I > MDD (P = 0.049); HCs > MDD (P < 0.001)
Duration (months)	44.98 (50.95)	49.37 (52.78)	N/A	−0.34	64	0.73	N/A
HAMD score	21.84 (4.50)	21.53 (4.17)	N/A	0.29	64	0.78	N/A
YMRS score	1.68 (1.76)	4.08 (3.43)	N/A	−3.52	52	0.001	BDD-I < MDD (P = 0.001)
Cognitive function
WAIS—Information	19.16 (5.06)	17.92 (5.67)	20.92 (4.66)	4.73	2	0.01	HCs > MDD (P = 0.004)
WAIS—Digit	68.65 (12.94)	67.88 (16.60)	89.06 (13.08)	40.16	2	< 0.001	HCs > BDD-I (P < 0.001), HCs > MDD (P < 0.001)
Behavioral data of the n-back task
0-back ACC (%)	89.90 (15.01)	90.58 (14.87)	92.01 (14.71)	0.27	2	0.77	N/A
2-back ACC (%)	60.77 (28.28)	65.61 (18.11)	72.52 (19.32)	3.80	2	0.025	HCs > BDD-I (P = 0.039)
0-back RT (ms)	555.01 (117.85)	537.76 (118.66)	484.85 (90.21)	6.49	2	0.002	BDD-I > HCs (P = 0.005), MDD > HCs (P = 0.01)
2-back RT (ms)	777.05 (199.40)	701.02 (179.12)	649.72 (139.31)	6.91	2	0.001	BDD-I > HCs (P = 0.002)
Medication
Total DDD	4208.32	1041.67	N/A	2.38	64	0.02	N/A
Mood stabilizers (n)	14	1	N/A	16.75	1	< 0.001	N/A
Antidepressants (n)	21	23	N/A	0.03	1	0.86	N/A
Antipsychotics (n)	16	4	N/A	12.59	1	< 0.001	N/A
Benzodiazepines (n)	6	15	N/A	4.19	1	0.041	N/A
Unmedicated (n)	1	8	N/A	5.38	1	0.02	N/A

Note: BDD-I: depression in bipolar disorder type I; df: degrees of freedom; DDD, defined daily dose; HCs: healthy controls; HAMD: 17-item Hamilton Depression Rating Scale; MDD: major depressive disorder; N/A: not available; WAIS: Wechsler Adult Intelligence Scale; YMRS: Young Mania Rating Scale; 0-back ACC: accuracy of the target under the 0-back load; 2-back ACC: accuracy of the target under the 2-back load; 0-back RT: response time under the 0-back load; 2-back RT: response time under the 2-back load.

N-Back Task Performance

See Table 1 and Supplemental Figure S2 for accuracy and RT values. The performance accuracy was significantly higher for the 0-back than the 2-back load (91.22 ± 14.74% vs. 68.37 ± 21.66%), and patients with BDD-I had significantly lower accuracy than HCs for the 2-back load (t = -2.13, P = 0.039). The RT for the 0-back load was significantly shorter than the 2-back load (512.43 ± 107.57 ms vs. 689.05 ± 169.90 ms). Patients with BDD-I and MDD both had significantly longer RT than HCs during the 0-back load (t = 2.99, P = 0.005; t = 2.62, P= 0.01), and patients with BDD-I also had longer RT than HCs during the 2-back load (t = 3.26, P = 0.002).

Differences in DC Among Groups

As shown in Table 2 and Figure 1, during the low WM load condition (0-back), patients with BDD-I showed lower DC in the mPFC, the precentral gyrus (PreCG), the ACC, the cuneus (CUN), the calcarine (CAL), the precuneus (PCUN), and the lingual gyrus (LING) compared with HCs. Compared to HCs, patients with MDD had higher DC in the mPFC, the inferior temporal gyrus (ITG), and orbital part of the inferior frontal gyrus (ORBinf) and lower DC in the postcentral cortex (PoCG), the ACC, the CUN, the CAL, the middle temporal gyrus (MTG), and the PCUN. When compared with patients with MDD, patients with BDD-I had lower DC in the mPFC (part of the DMN) along with clusters in the ORBinf.

Table 2.

Brain Regions with Significant Differences in DC Maps During n-Back Task.

Task	Comparisons	Brain region		Cluster	Peak coordinates			t
					x	y	z
0-back	BDD-I > HCs	N/A
	BDD-I < HCs	LING.R		24	9	−51	3	−4.40
		PCUN.R		18	3	−60	51	−4.37
		CAL.L		33	−12	−63	6	−3.95
		PreCG.R		6	24	−18	63	−3.40
		CUN.L		19	−6	−78	30	−3.13
		mPFC.R		5	24	51	21	−2.84
		ACC.L		8	0	39	21	−2.81
	MDD > HCs	mPFC.R		43	30	42	−18	4.27
		ORBinf.L		20	−15	12	−21	3.68
		ITG.L		7	−57	−51	−12	2.98
	MDD < HCs	MTG.L		10	−54	−51	9	−4.34
		PoCG.R		6	36	−33	63	−3.98
		CUN.L		19	−6	−75	30	−3.83
		CAL.L		12	−15	−75	9	−3.51
		PCUN.R		14	3	−60	48	−3.37
		ACC.R		9	3	30	18	−3.29
	BDD-I > MDD	N/A
	BDD-I < MDD	ORBinf.L		14	−18	18	−21	−3.77
		mPFC.R		17	27	42	−18	−3.05
2-back	BDD-I > HCs	IPL.L		7	−42	−36	36	3.36
	BDD-I < HCs	PreCG.R		35	27	−21	60	−4.48
		CUN.R		11	6	−90	30	−3.97
		mPFC.L		10	−42	9	51	−3.66
		DLPFC.L		5	−21	0	66	−3.64
		PCL.R		7	3	−24	69	−3.58
		PoCG.R		12	45	−18	51	−3.54
		CER.R		5	21	−51	−57	−3.32
		MTG.R		9	69	−18	−9	−3.20
		HIP.L		9	−21	−6	−21	−2.98
	MDD > HCs	mPFC.R		5	30	45	−18	3.32
	MDD < HCs	HIP.L		15	−18	−12	−18	−4.26
		PHG.R		5	18	−3	−24	−3.48
		THA.L		20	−18	−12	3	−3.34
	BDD-I > MDD	THA.L		8	−21	−15	0	3.67
		IPL.L		6	−42	−36	36	3.07
	BDD-I < MDD	mPFC.L		10	−39	9	54	−4.09
		CER.R		5	18	−54	−57	−3.62
		MTG.R		7	66	−15	−9	−3.60
		PreCG.R		16	30	−15	60	−3.51
		mPFC.R		5	27	48	−18	−3.04

Note: P < 0.05, FDR correction. ACC: anterior cingulate gyrus; BDD-I: depression in bipolar disorder type I; CAL: calcarine; CER: cerebellum; CUN: cuneus; DC: degree centrality; DLPFC: dorsolateral prefrontal cortex; HCs: healthy controls; HIP: hippocampus; ITG: inferior temporal gyrus; IPL: inferior parietal cortex; MDD: major depressive disorder; L: left; LING: lingual gyrus; mPFC: medial prefrontal cortex; MTG: middle temporal gyrus; N/A: not available; ORBinf: orbital part of Inferior frontal gyrus; PCL: paracentral lobule; PHG: paraHippocampal gyrus; PCUN: precuneus; PreCG: precentral gyrus; PoCG: postcentral cortex; R: right; THA: thalamus.

Figure 1.

Statistical DC map under 0-back working memory (WM) load. Results from the between-groups comparison (t-tests) showing clusters with significant differences in DC values. The t-maps of DC values (t-tests) were generated at FDR-corrected P < 0.05. Note: BDD-I: depression in bipolar disorder type I; DC: degree centrality; HCs: healthy controls; MDD: major depressive disorder.

As shown in Table 2 and Figure 2, during the high WM load condition (2-back), patients with BDD-I had higher DC in the inferior parietal cortex (IPL) and lower DC in the mPFC, the PreCG, the PoCG, the DLPFC, the paracentral lobule (PCL), the MTG, the hippocampus (HIP), the CUN and the cerebellum than HCs. Compared to HCs, patients with MDD had higher DC in the mPFC and lower DC in the parahippocampal gyrus (PHG), the HIP, and the thalamus (THA). When compared with patients with MDD, patients with BDD-I had higher DC in the IPL and the THA, and lower DC in the mPFC (part of the DMN), the PreCG (part of the sensorimotor network [SMN]), the MTG, and the CER.

Figure 2.

Statistical DC map under 2-back working memory (WM) load. Results from the between-groups comparison (t-tests) showing clusters with significant differences in DC values. The t-maps of DC values (t-tests) were generated at FDR-corrected P < 0.05. Note: BDD-I: depression in bipolar disorder type I; DC: degree centrality; HCs: healthy controls; MDD: major depressive disorder.

We used other thresholds (r = 0.15 and r = 0.35) to validate our results and found that the choice of these thresholds did not significantly affect the main results, and these results are shown in Supplemental Table S5.

The significant omnibus differences of the DC map among the three groups during the n-back task (Supplemental Table S6) and resting periods of the n-back (Supplemental Table S7 and Figure S3) are shown in the Supplemental Material. In brief, during resting periods of the n-back task, patients with BDD-I showed lower DC in the mPFC and the cerebellum relative to patients with MDD. We also analyzed the resting-state fMRI data (collected separately from the n-back task, but in the same scanning session) for completeness. The results of ANOVA and post hoc t-tests of DC map among the three groups during resting-state are shown in Supplemental Material S4 and Tables S1–S2 and the results of global signal regression are shown in Supplemental Tables S3–S4.

Exploratory Analysis

Correlation analysis: No statistically significant correlations were found between altered DC value in patients and medication, clinical characteristics, cognitive function, and WM performance (see Supplemental Material S6).

Machine learning analysis: We retrieved the DC value of the regions that showed significant differences between BDD-I and MDD as features to conduct the subsequent classification analyses. There were 2, 2, 7 features for the classification analyses for the resting periods, 0-back load, and 2-back load, respectively. We achieved classification results with average accuracy, the true positive rate for BDD-I, the true-positive rate for MDD based on the features retrieved from the resting periods of n-back as 64.81%, 0.64, 56%, and 72.41%, respectively; from the 0-back load as 70.49%, 0.7%, 64.3%, and 75.8%, respectively; from the 2-back load as 90.91%, 0.91%, 92.3%, and 89.7%, respectively. With the increase in the task load, the accuracy and AUC for classifying MD from BDD-I increased in this sample. The receiver operating characteristics are shown in Figure 3.

Figure 3.

The receiver operating characteristics of the machine learning analysis. The AUC under resting periods, 0-back, and 2-back load is 0.64, 0.7, and 0.91, respectively. Note: BDD-I: depression in bipolar disorder type I; MDD: major depressive disorder.

Discussion

Patients with mood disorders often suffer from cognitive dysfunction. ⁹ Here, we examined the differences in the connectivity of brain functional networks during a cognitive (WM) task in depressed patients with BDD-I and MDD. The main findings were as follows: When task load was increased, we were able to classify BDD-I from MDD with a notable gain in classification accuracy using the DC maps. The distinct DC map of 2-back load produced a within-sample true positivity rate of 92.3% for BDD-I. This indicates that a signature pattern for BDD-I in the functional connectome can be seen when a subject is placed under a degree of cognitive demand. Patients with BDD-I displayed decreased DC in the DMN (the mPFC) and the SMN (the PreCG and the PoCG) during high WM load conditions. In contrast, patients with MDD displayed increased DC in the DMN (the mPFC) during high WM load conditions.

The key finding of the present study was that the patients with BDD-I showed a lower DC value (a measure of both local and distant connectivity) affecting the DMN and the SMN compared to patients in the MDD group and HCs during a high WM load condition. These findings are in line with the concentrated changes in SMN during a phase of depression as reported by Russo and colleagues ³¹ and the DMN/SMN reported by Martino and colleagues. ³² The role of the DMN has been well explored in MDD, especially in relation to cognitive aspects of depression. ³³ With respect to the SMN, the PreCG consists of the primary motor cortex that is involved in different aspects of motor planning and decision making. ³⁴ This network is seen as a key player in psychomotor aberrations in depression. ³⁵ Several other resting-state fMRI studies have reported alterations in FC of SMN in bipolar disorder regardless of the phase of illness of patients.^36,37 To our knowledge, our observation is the first report of reduced DC value during WM task in patients with BDD-I, and this abnormal DC value on the SMN in BDD might be accountable for their impaired cognitive function. While there are methodological differences including the task-fMRI focus and depressive phase in this study, and the lack of band-specific examination of fluctuations, our observations extend support for the postulated role for DMN and SMN in both BDD-I and MDD. ³⁸

Further, patients in the MDD group showed increased DC values compared to patients in the BDD-I group and HCs in the DMN at high WM load. Connectivity within the DMN and between the DMN and other brain regions during resting-state was frequently reported to be increased in the MDD group, and researchers have pointed out that the increased DMN connectivity might be related to negative rumination in patients with MDD.^39,40 In addition, many researchers have found increased DMN activation during a WM task in patients with MDD and indicated the reduced task-induced deactivation of the DMN might be a source of cognitive dysfunction in this population.^17,41 Scalabrini and colleagues highlighted the role of DMN abnormalities in the whole-brain connectivity in MDD. ⁴² Moreover, the DMN is involved in internal mental states and correspondent internally driven behavior, including ruminations and increased self-focus which are closely associated with depressive symptomatology,^43–45 and it also plays a core role in using past experiences to plan for the future, evaluating survival cues in the external environment and maximizing the utility of moments when we are not distracted by the external world.^45–47 Hence, the increased DC value during task performance in patients with MDD might be interpreted as the patients’ inability to navigate away from internal emotional and cognitive states (e.g., rumination), which is prohibiting them from focusing on the current task.

We retrieved the DC value of the regions that show significant differences between BDD-I and MDD under three conditions (i.e., resting periods, 0-back load, and 2-back load) as features to conduct the classification analyses separately. We found that the higher WM task load was accompanied by a larger difference between groups (superior discriminatory accuracy). Wherein, the accuracy under 2-back load is 90.91%, which is higher than the clinically useful threshold of 80%. ⁴⁸ It seems that when the load for working memory maintenance is higher (i.e. 2-back, compared to 0-back), the connectome profile of the two groups of depressed patients become more distinguishable from each other. It should be noted that our accuracy under resting period (64.8%) is lower than a prior study reported accuracy (86%) under resting-state, ²² and we used the same neuroimaging feature of DC. We speculate that this discrepancy may relate to differences in the sample, as well as acquisition differences. We used a shorter acquisition time for our rest block (80 volumes) compared to the prior study that employed a longer time course (240 volumes). ²² Further evaluations are required to determine the ideal acquisition time for resting-state scans to achieve optimal discriminatory accuracy.

Some limitations must be considered when appraising our observations. First, the sample size in our study is relatively small, though it is sufficient to demonstrate task-related connectivity changes. Further, patients in our study were receiving psychotropic medications, but correlation analysis of the effect of drug dosages on the altered DC value in patients produced no notable associations. To evaluate the real-world clinical utility of any test that purports to differentiate BDD-I and MDD, the ideal study design would be to recruit many subjects with first-episode depression, administer the WM task in an fMRI, and follow them over several years to estimate the risk of conversion to bipolar disorder from MDD. While depression is the index episode for a high proportion of BDD (50%–80%), ⁴⁹ longitudinal studies estimate an average 6–8 years for an episode of mania to occur after the first depressive episode,^50,51 with the risk of conversion to be <10% over this period even when a large population database is employed. ⁵² Thus, employing a longitudinal design to evaluate the discriminatory utility of a WM task using fMRI in a first-episode depressed cohort is a formidable challenge. Nevertheless, enriching first-episode depression samples based on clinical and genetic indicators of bipolarity (see O’Donovan & Alda for further discussion), ⁴⁹ or identifying cases of “missed diagnosis” through survey approaches, ⁵³ are more practical next steps within the evidentiary framework. Our study is best considered as an initial “candidate identification” approach. Our assumption that the task-related functional connectome changes seen in established cases of BDD-I have predictive value for discrimination at first presentation needs to be tested in future studies.

Conclusions

We report distinct patterns of connectome configuration in patients with BDD-I and MDD when processing working memory load. When examined using DC maps, this pattern might provide evidence to separate depression in bipolar disorder from unipolar depression.