Authors: I. Tahir, A. Planat-Chrétien, A. Bertrand, M. Linder, C. Dondé, R. Sosnik, M. Polosan
Categories: Article, Diagnostic markers, Bipolar disorder
Source: Translational Psychiatry
Authors: I. Tahir, A. Planat-Chrétien, A. Bertrand, M. Linder, C. Dondé, R. Sosnik, M. Polosan
Bipolar disorder (BD) is a complex mood disorder characterized by recurrent depressive and manic/hypomanic episodes, accompanied by significant cognitive dysfunction and emotional dysregulation. Accurate and timely diagnosis, especially the differentiation between subtypes, remains a challenge due to overlapping symptoms, variable onset times for more specific symptoms (e.g., psychotic features), and the reliance on subjective assessments. This study examines the use of a multimodal approach combining electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) to identify patterns of BD emotional dysregulation, aiming to enhance its diagnosis and subtype differentiation. The protocol employed an emotional visual task to evaluate the interference of emotional content on cognitive function. EEG data were collected using a whole-head cap, while fNIRS focused on hemodynamic changes in the frontal cortex. Furthermore, the feasibility of using a potential simplified, portable EEG–fNIRS system was explored by focusing the analysis on frontal regions. The cohort included BD patients [BP] of two main subtypes, and healthy controls [HC]. Behavioral analysis revealed significant performance differences between BP and HC groups. While EEG alone enabled groups’ classification, integrating EEG and fNIRS improved accuracy by reducing misclassification rates. Although classification using only frontal EEG regions was slightly less accurate than the full-head cap, fNIRS integration ensured robust results, supporting the feasibility for a potential simplified system. These findings underscore the complementary strengths of EEG and fNIRS in capturing neural and vascular markers of emotional dysregulation in BD and support the development of multimodal diagnostic tools for BD.
Bipolar disorder (BD) is a mood disorder affecting over 1% of the population, characterized by alternating episodes of depression and mania/hypomania, interspersed with intercritical, euthymic periods. BD is classified into two primary type I, involving at least one manic episode, and type II, characterized by at least one major depressive episode and one hypomanic episode. The diagnosis relies solely on clinical assessment, as there are no complementary diagnostic tests. Early and accurate diagnosis is crucial for optimizing treatment, which may vary depending on subtype (i.e. type I patients often respond better to lithium than to other medications [1]), highlighting the importance of distinguishing subtypes. Timely diagnosis also aids in reducing comorbidities and complications, including suicide.
BD patients (BP) experience mood swings that may stem from emotional dysregulation, persisting during euthymic periods [2–4]. This dysregulation impairs cognitive performance on tasks involving emotional interference - a slow-down of information processing due to emotional valence of presented stimuli, and is associated with dysfunctions in the brain’s emotional regulation network. It is characterized by imbalanced interactions between frontal regions, which govern voluntary and automatic emotional regulation, and limbic regions [5]. The neuropathophysiological basis of the disorder is becoming clearer due to advances in neuroimaging techniques. Functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) have revealed dysconnectivity between frontal and limbic areas, including hyperactivation of the amygdala and ventral striatum, hypoactivation of the frontal cortex [6–9], and disruptions in white matter tracts linking frontal and limbic regions [10–12]. While these techniques offer valuable diagnostic and therapeutic insights, their high cost, complexity, low temporal resolution, and lack of portability hinder their widespread adoption in routine clinical practice and integration with more complex tasks.
Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) are non-invasive methods suitable for non-hospital settings use, allowing for more realistic protocols. They measure signals from superficial cortical areas such as the frontal cortex, a part of the emotional regulation network, potentially enabling the identification of emotional dysregulation markers associated with BD. The placement of fNIRS on the frontal region is advantageous enabling the measurement of key cortical functions related to decision-making, emotional regulation, and social interactions [13].
Integrating EEG and fNIRS leverages their complementary strengths, with fNIRS providing good local spatial resolution, EEG ensuring precise temporal resolution, and their combined features enabling the neurovascular coupling analysis [14]. The combination of these modalities is particularly relevant in BD, as neuroinflammatory factors can affect the blood-brain barrier, potentially contributing to the cognitive impairments found in BP [15]. In this context, Trambaiolli et al. [16] introduced the concept of EEG spectro-temporal amplitude modulation (EEG-AM) as a feature correlated with fNIRS-derived total hemoglobin concentration, showing that EEG-AM carries hemodynamic information. More recently, Lin et al. [17] proposed a bimodal EEG–fNIRS framework to study neurovascular coupling during cognitive–motor interference, reporting reduced coupling in theta, alpha, and beta bands and underscoring the complementary value of multimodal integration. Building on this, Kopf et al. [18] exploited EEG–fNIRS system during an emotional n-back paradigm to analyze how emotional interference affects working memory (cognitive task) of BP, targeting the emotion regulation networks. They revealed altered frontal cortical oxygenation patterns and Late Positive Potential (LPP) dynamic.
The current study employed a system combining a full-head EEG cap and fNIRS focused on frontal regions during an emotional visual task, aiming to assess the impact of emotional interference on cognitive function. We hypothesize that this approach will facilitate the identification of biomarkers of emotional dysregulation in BD through cortical measures, thereby aiding in the diagnosis of the disorder. Moreover, given that bipolar subtypes exhibit distinct clinical profiles and are likely driven by different neuropathophysiological mechanisms [19, 20], we speculate that they can be discriminated using neurovascular features.
The primary objectives of the study (1) to investigate whether emotional dysregulation could be studied via cortical structures activity to classify groups (BP, healthy controls [HC], and BD subtypes); (2) to determine whether the integration of EEG and fNIRS improves the classification compared to using EEG alone; and (3), to explore the feasibility of groups classification using a potential simplified, portable EEG–fNIRS system covering only frontal regions.
The study cohort consisted of 25 HC (10 males,15 females, 32.4 ± 14.8 years old) and 46 BP, including 21 type I (BP I) (7 males,14 females, 50.1 ± 12.4 years old) and 25 type II (BP II) (15 males,10 females, 48.4 ± 12.0 years old). The clinical severity was assessed using standardized the Clinical Global Impression (CGI) scale [21], the Montgomery–Åsberg Depression Rating Scale (MADRS) [22], the Young Mania Rating Scale (YMRS) [23] and the Global Assessment of Functioning (GAF) scale [24] (see Supplementary Table 1 for further details of participants’ characteristics).
Inclusion criteria for HC and BP required participants to be between 18 and 70 years old, with no history of alexithymia, and no history of neuropsychiatric disorders for HC. BP needed to meet the criteria for either type I or II as defined by the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), which also allows for the distinction between depressive/hyperthymic episodes and intercritical euthymic period. Exclusion criteria for BP and HC included women with Premenstrual Dysphoric Disorder, individuals with color blindness or weakness, and participants affected by drug or alcohol use. The study was approved by the local ethics committee (CPP Sud Est II - 2021-A00290-41, France) and adhered to the principles of the Declaration of Helsinki (NCT05025566). All participants were provided with study information, and written informed consent was obtained prior to participation. Data with technical issues or incomplete datasets were excluded. As a result, the cohort analyzed consisted of 18 HC and 40 BP (19 BP I and 21 BP II).
The study employed an emotional visual task [25], based on the emotional Stroop task designed by Quan et al. [26]. Emotional images interfering with a simple cognitive task were used, as shown in Fig. 1. The task was designed using E-Prime 3 (Psychology Software Tools, Pittsburgh, PA). Participants first completed a resting period (eyes closed), followed by a training block of 16 neutral images. Next, they performed a task comprising six blocks of images with different emotional two blocks with neutral images, two blocks with negative images, and two blocks with positive images. Each block comprised 20 trials, with a 50-second rest period between blocks and randomized inter-stimulus intervals (ISI) of 9 to 11 seconds. In each block, a given image appeared twice - surrounded by a green or a red frame. The order of image presentation was randomized for each participant. During each trial, participants were instructed to respond ‘as quickly and as accurately as possible’ to the color of the frame surrounding the image by pressing the corresponding keyboard key, while ignoring the emotional content of the image. The timing of stimuli presentation and responses (correct, incorrect, or no response) were recorded for offline analysis. The emotional images were selected from the Open Affective Standardized Image Set (OASIS) [27], a database with normalized images (brightness, color, and visual complexity), and included humans, animals, objects, and scenes. The mean valence and arousal of the selected images negative (2.86 ± 0.94; 3.97 ± 0.54), neutral (4.43 ± 0.61; 2.42 ± 0.50), positive (5.80 ± 0.36; 3.88 ± 0.59).Fig. 1Protocol design.Participants were instructed to respond ‘as quickly and as accurately as possible’ to the color of the frame surrounding the image, while ignoring the emotional content of the image, by pressing the corresponding keyboard key. Yellow arrow - message “The experiment begins”. Green arrow – message “The experiment continues”. Blue arrow – message “The experiment ended”. The order of the six experimental blocks (two neutral, two negative, and two positive) as well as the selection of pictures within each block were randomized.
The bimodal system, presented in Fig. 2, parallel recorded EEG data and fNIRS data on two systems synchronized via a common trigger. EEG data were recorded using a 64-channel active system (ActiCap 64ch, electrode positions conformed to 10-20 system) with BrainVision Analyzer software. The reference electrode for EEG was placed in front of the Cz, and the ground electrode on the forehead above the nose. Electrooculogram (EOG) signals were recorded using four two for vertical EOG (above and below the left eye) and two for horizontal EOG (at the outer canthus of the eyes). EEG data were recorded at a sampling rate of 1 kHz. The fNIRS setup (OXYMON MK III, Artinis) included four sources and four detectors, defining six channels, as shown in Fig. 2. Due to hardware limitations, optodes were restricted to the left frontal cortex, consistent with previous findings reported in the literature [28–31]. The inter-optode separation was 4 cm, and two wavelengths (760 nm and 850 nm) were employed to capture changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) concentrations.Fig. 2Data acquisition system.(a) The system including the 64-channel full-head electroencephalography (EEG) cap conformed to the international 10–20 system and frontal functional near-infrared (fNIRS) sensors with four sources and four detectors, defining a total of 6 fNIRS channels (ch1 to ch6). (b) Placement of electrodes and optodes.
For demographic variables, sex differences were tested with a Chi² test (global) and Fisher’s exact tests (pairwise, effect size reported as odds ratio), while medication use (antidepressants AD, antipsychotics AP, mood stabilizers MS, and anxiolytics ANX) was compared between BP I and BP II using Fisher’s tests (effect size reported as odds ratio). Depending on normality (Shapiro–Wilk test), either Wilcoxon rank-sum tests (with effect size r) were applied for non-normal groups (BP I vs. HC, BP II vs. HC), or a t-test (with Hedges’ g as effect size) was used for BP I vs. BP II.
In order to assess the effect of the condition (image valence) on cognitive performance (recognition of the color of the frame surrounding the image), response time and number of correct responses (accuracy) were computed. First, within-group statistical tests were carried out to evaluate the effect of the color frame (Red vs. Green) across the Negative, Neutral, and Positive conditions on performance. Depending on normality, either a paired t-test (with Hedges’ g as effect size) or the Wilcoxon signed-rank test (with effect size r) was applied. Pairwise comparisons between emotional conditions (Negative vs. Neutral, Negative vs. Positive, Neutral vs. Positive) were conducted separately for each group (BP, HC, BP I, BP II) using the same tests. Subsequently, for each condition (Neutral, Negative, Positive) the non-parametric Kruskal–Wallis test was employed to compare groups’ medians— HC, BP I, and BP II. If a significant difference was found, a Wilcoxon rank-sum test with Holm correction for multiple testing was used as a post-hoc test to perform pairwise groups comparisons. Effect sizes r were computed for all Wilcoxon tests.
The following pipeline was implemented using MATLAB 2022B toolboxes.
To mitigate EEG artefacts, a band-pass filter between 1–45 Hz was used. Next, a home-made adapted version of Artefact Subspace Reconstruction (ASR) [32] was used to remove eye blinks, muscle movements, and non-biological artifacts in recordings.
A band-pass filter between 0.01 – 0.3 Hz was used to remove low-frequency noise and physiological artifacts, including heartbeats. Motion artifacts were corrected using Tukey’s Biweight Robust Mean method [33]. To further minimize extracerebral contributions, Principal Component Analysis (PCA) was performed, and only the first principal component was removed, based on the assumption that it primarily captures extracerebral and systemic physiological noise [34]. The concentration changes in HbO and HbR are estimated from these corrected signals at two wavelengths based on the classical Modified Beer–Lambert Law approach [35].
Specific EEG and fNIRS features were extracted from preprocessed data. EEG signals were segmented into 2-second epochs aligned with the onset of each stimulus. Event-related potentials (ERP) were computed for each subject, by averaging these signals for each stimulus condition and across specific brain regions frontal (AF3, AF4, Fz), parietal (CP1, CP2, P3, P4, Pz), central (FC1, FC2, C3, C4, Cz), and occipital (PO3, PO4, O1, O2, Oz) areas.
Concerning fNIRS, concentration signals were segmented into 10-s epochs aligned with stimulus onset in order to capture the full hemodynamic response. Although stimuli lasted only 2 s, the use of a ~ 10-s ISI [36, 37] ensured complete hemodynamic recovery, prevented trial-to-trial overlap [38], and minimized potential residual emotional effects. Stimulus-evoked hemodynamic responses (SHR) were then obtained for each subject in HbO and HbR by averaging the signals across trials of each condition.
Integrals were then computed for ERP and SHR within specific time windows of interest, resulting in two feature vectors (EEG, fNIRS) that were subsequently processed and used for classification purposes. Time windows were defined based on a preliminary grand average analysis for both modalities, as shown in Fig. 3. For EEG, a sliding-window analysis with a 70-ms step size was applied to the grand-average ERP time courses (e.g., grand average over the occipital region in the HC group, shown in Fig. 3a) for each condition. Group comparisons were conducted separately for HC vs. BP I, HC vs. BP II, and BP I vs. BP II using two-sided Wilcoxon rank-sum tests (unpaired). To address multiple testing across windows, p-values were adjusted using the Benjamini–Hochberg false discovery rate (FDR) per condition and group comparison. This procedure was used solely to identify the time intervals for subsequent feature extraction. Finally, four time windows were defined, centered around peaks corresponding to various cognitive processes, in accordance with literature [39–41]: a peak associated with visual processing at 100 ms (70-130 ms), an attention-related peak at 250 ms (200-300 ms), an emotion-processing peak at 450 ms (450-650 ms), and a late peak at 1000 ms (900-1100 ms) linked to higher-level cognitive processes, decision-making and emotional process.Fig. 3Grand average ERP and fNIRS signals with identified time windows.(a) Example of a grand average event-related potential (ERP) signal recorded over the occipital area (electroencephalography - EEG), highlighting the time windows selected based on significant intergroup differences, identified using Wilcoxon rank-sum tests and supported by existing literature. (b) Example of grand average functional near-infrared (fNIRS) signals showing changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) concentrations. The three selected time windows—2–4 s, 4–6 s, and 6–8 s—are marked and used for subsequent analysis.
For fNIRS, three time windows were defined based on grand average 2-4 s, 4-6 s, 6-8 s, as illustrated in Fig. 3b. For each window, the integral of the HbO and HbR signals was computed. These three windows provide insight into the typical delay of the hemodynamic response, the phase of maximum hemodynamic activity, and the late recovery phase.
Finally, 216 fNIRS features were extracted (six conditions, six channels, two hemoglobin types, and three-time windows), while 96 EEG features were computed (six conditions, four brain regions, and four-time windows).
To leverage the complementary nature of the two modalities, a feature fusion approach was implemented. While a whole-head EEG cap covers more brain regions, fNIRS measurements were limited to the frontal area. Consequently, concatenating EEG and fNIRS features tends to favor the selection of EEG features that capture connectivity patterns across regions. To address this imbalance, a fusion approach was applied, combining the most relevant normalized features selected separately for each modality.
To reduce dimensionality while retaining the most discriminative information, we implemented a sequential forward feature selection strategy applied separately to EEG and fNIRS features. The objective function was defined as the minimization of the cross-validated classification at each step, a support vector machine (SVM) with a radial basis function (RBF) kernel was trained, and the objective function returned the number of misclassified samples in the validation folds. The procedure was embedded in a loop of 50 runs of stratified 5-fold cross-validation, ensuring that feature selection was performed in a cross-validated setting. At the end of each run, sequential forward selection identified a subset of features that minimized classification error and thus contributed most to performance. A recurrence analysis was performed across the 50 runs, and only features consistently present in more than 30% of the runs (heuristic threshold) were retained, yielding a stable and discriminative feature subset for each modality. Finally, the selected EEG and fNIRS features were fused into a multimodal feature vector, which was used as input for subsequent group classification. To ensure robustness and avoid overfitting, the maximum number of selected features was defined according to best practices for small cohorts [42]. This number has actually never been reached in our analysis.
Group classification was carried out using a Leave-One-Subject-Out (LOSO) cross-validation framework, in which the data from N–1 subjects were used for training and the remaining subject was held out for testing. This subject-level validation ensured complete independence between training and testing sets, preventing any data leakage. Within each LOSO fold, hyperparameters of the SVM classifier (kernel linear, polynomial, radial basis function [RBF], Gaussian; box-constraint C ∈ {0.1, 1, 10, 100}) were optimized exclusively within the training set using a nested stratified 5-fold cross-validation. After optimal parameters were identified, the model was retrained on the full training set and evaluated on the held-out subject. This process was repeated until each subject had served once as the test case. After completing all folds, predictions were aggregated to construct a confusion matrix at the subject level. From this raw matrix, we computed accuracy, sensitivity, specificity, precision, and F1-score. The area under the curve (AUC) was calculated from the per-subject decision scores. All metrics were further reported with 95% confidence intervals across subjects. Confusion matrix was row-normalized by the number of true subjects in each class.
To address class imbalance in BP and HC group classification, random subsampling permutations were performed, each time generating balanced subsets that were subsequently used for classification. All metrics were reported with 95% confidence intervals across random balanced permutations.
As a first result, this pipeline was used to compare classification performance between EEG–fNIRS combination and EEG alone, using whole-head EEG. Then, to evaluate the feasibility of a simplified system, the analysis was repeated using the same EEG–fNIRS system, with EEG electrodes restricted to the frontal regions. The classification performance was re-evaluated to assess its robustness.
Regarding demographic variables, no significant differences in sex distribution were observed across groups (global Chi², p = 0.25, V = 0.28, 95% CI [0.00, 0.43]; all pairwise Fisher’s tests, p > 0.05). Similarly, no significant differences were found in medication use between BP I and BP II patients across all drug classes (AD, AP, MS, ANX; all p > 0.05). In contrast, age differed significantly between patients and healthy controls. Both BP I and BP II were older than HC (Wilcoxon tests, p = 0.0002, r = 0.55, 95% CI [0.31, 0.74] and p = 0.0004, r = 0.50, 95% CI [0.24, 0.72] respectively), with large effect sizes. No age difference was observed between BP I and BP II (t-test, p = 0.64, g = 0.14, 95% CI [−0.44, 0.72]).
No statistically significant differences in response times or accuracy were found between images with a red frame and those with a green frame (all p > 0.05). Similarly, no significant differences in accuracy were observed between groups (all p > 0.05). In line with the literature [26], pairwise analyses revealed significant differences between Negative and Neutral conditions in BP (p = 0.003, g = 0.15, 95% CI [0.05 0.25]), present in both BP I (p = 0.04, g = 0.22, 95% CI [0.02 0.43]) and BP II (p = 0.04, g = 0.11, 95% CI [0.006 0.216]), as illustrated in Fig. 4a. Between-group comparisons, presented in Fig. 4b, further showed significantly longer response times in BP I compared to HC across all conditions (all p-values < 0.05), whereas BP II did not differ from HC (all p > 0.05). However, BP II showed significant differences in response times across emotional conditions, whereas HC did not exhibit any condition-dependent variations (all p > 0.05). Detailed statistical results for behavioral and demographical analyses are provided in Supplementary Tables 2–8.Fig. 4Behavioral performance.(a) Comparison of response time across emotional visual stimuli (neutral, positive, and negative valence pictures) for each group. (b) Comparison of response time between groups for each condition. Patients with bipolar disorder (BP), including BP type I (BP I) and BP type II (BP II), and healthy controls (HC). P-values are indicated as * indicates significance at p < 0.05; ** indicates significance at p < 0.01.
Using EEG alone with a whole-head montage already provides good classification performance across the groups - BP vs. HC (Fig. 5a) were effectively distinguished with an accuracy of 0.78 (95% CI [0.76–0.79]), increasing to 0.79 (95% CI [0.78–0.80]) with the addition of fNIRS features. The BP I vs. HC (Fig. 5b), BP II vs. HC (Fig. 5c), and BP I vs. BP II (Fig. 5d) comparisons yielded accuracies of 0.86 (95% CI [0.72–0.94]), 0.77 (95% CI [0.62–0.87]), and 0.80 (95% CI [0.65–0.89]), respectively, when using EEG only. When fNIRS signals were added to EEG, the accuracy increased to 0.92 (95% CI [0.79–0.97]) for BP I vs. HC, 0.85 (95% CI [0.70–0.93]) for BP II vs. HC, and 0.85 (95% CI [0.71–0.92]) for BP I vs. BP II. These improvements were accompanied by systematic gains in performances metrics—including specificity, sensitivity, precision, F1-score, and AUC—reported in Supplementary Table 9.Fig. 5Classification accuracy using whole-head EEG montage.Confusion matrices showing the classification performance for (a) patients with bipolar disorder (BP) vs. healthy controls (HC), (b) BP type I (BP I) vs. HC, (c) BP type II (BP II) vs. HC and (d) BP I vs. BP II. For each plot, the classification accuracy using electroencephalography (EEG) alone (left) and EEG combined with functional near-infrared (fNIRS) (right) is presented.
Relevant EEG features across comparisons are described in Supplementary Table 10, which summarize the top-ranked features retained during the feature selection process for each group classification. We noticed that frontal EEG features were identified as relevant for group classification using the whole-head montage. This finding led to the decision to focus the EEG analysis on this area and assess the feasibility of a simplified, frontal-focused system.
Restricting EEG analysis to frontal regions provides lower classification performance compared to the whole-head montage. BP vs. HC were distinguished with an accuracy of 0.72 (95% CI [0.70–0.73], Fig. 6a), which increased to 0.74 (95% CI [0.72–0.75]) when fNIRS features were added. For BP I vs. HC, EEG alone reached a good accuracy of 0.84 (95% CI [0.69–0.92]), but performance slightly decreased with the bimodal approach to 0.81 (95% CI [0.66–0.90]), although it remained high overall.Fig. 6Classification accuracy using frontal regions.The confusion matrices show the classification performance for (a) patients with bipolar disorder (BP) vs. healthy controls (HC) (top row, left column), (b) BP type I (BP I) vs. HC, (c) BP type II (BP II) vs. HC and (d) BP I vs. BP II. In each plot, the classification accuracy using electroencephalography (EEG) alone (left) vs. EEG combined with functional near-infrared (fNIRS) (right) is presented.
In contrast, the BP II vs. HC and BP I vs. BP II comparisons yielded lower accuracies of 0.67 (95% CI [0.51–0.79]) and 0.70 (95% CI [0.55–0.82]), respectively, for EEG only. When fNIRS features were added to frontal EEG, substantial improvements were the accuracy for BP II vs. HC increased to 0.82 (95% CI [0.67–0.91]), and the accuracy for BP I vs. BP II increased to 0.75 (95% CI [0.60–0.86]).
These improvements were accompanied by a general enhancement across all performance metrics, provided in Supplementary Table 11, leading to a more favorable overall balance.
As with the whole-head analyses, the features most often retained are reported in the Supplementary Table 12.
Accurate diagnosis of BD and its subtypes is essential due to their differing treatment approaches and clinical management. However, clinical assessment alone has proven insufficient for differentiation of the subtypes. For example, the occurrence of a manic episode in a patient previously diagnosed with BP II reveals a diagnostic misclassification, indicating that the correct diagnosis is BP I. This underscores the need for objective biomarkers of BD and its subtypes.
In our study, the behavioral analysis revealed that BP exhibited significantly longer response times to negative stimuli compared to positive/neutral ones, suggesting a heightened sensitivity to emotional valence compared to HC. Classification results using a whole-head montage highlighted that although EEG alone constitutes a relevant biomarker, the EEG–fNIRS system emerges as a promising clinical tool for BD diagnosis. In particular, the addition of fNIRS playing a crucial role in distinguishing BP II from HC by providing valuable neurovascular insights. Thus, this bimodal system could effectively differentiate BP from HC and further refine the distinction between subtypes through specific neuropathophysiological markers.
Regarding classification using frontal regions - EEG alone, consistently showed lower performance compared to whole-head EEG. However, the contribution of fNIRS provided perfusion-related information, complementing EEG data, helping reduce misclassification rates and ensuring more reliable results, especially when EEG spatial coverage was limited. Notably, by capturing subtle neurovascular differences, fNIRS improves the ability to differentiate BD subtypes and distinguish BP II from HC.
Indeed, previous studies highlighted that the clinical distinctions between BP I and BP II are paralleled by neurobiological and neurovascular differences. These include white matter microstructure, neurochemical concentrations, cortical morphology, and vascular function [43–45].
Recognizing and understanding these differences may inform more tailored treatment strategies and improve diagnostic specificity. Therefore, our findings pave the way for a simplified system able to distinguish the neurobiological and neurovascular differences underlying the clinical BD subtypes. Beyond BD, this accessible system could be an additional tool to clinical assessment for differentiating other mood disorders with overlapping symptoms, such as major depressive disorder (MDD) [46–48]. This study may also represent a preliminary step toward the development of a method enabling an early diagnosis of BD, which is known to reduce the duration of untreated illness—a factor closely linked to poorer prognosis [49]. Moreover, EEG–fNIRS method may help the identification of biomarkers related to individuals at risk and thus facilitate the development of preventive strategies and enable earlier interventions.
Recent studies have highlighted the value of fNIRS in mood disorder research. Mao et al. [50] examined prefrontal hemodynamic responses in BP during emotional induction and verbal fluency tasks, revealing distinct activation patterns compared to healthy controls. Tran et al. [51] demonstrated the diagnostic potential of portable fNIRS systems in unipolar and bipolar depression, showing that frontal hemodynamic markers—particularly during verbal fluency—can achieve clinically relevant classification accuracy. Feng et al. [52] further identified the ventrolateral prefrontal cortex as a discriminative region when comparing activation patterns between unipolar and bipolar depression. These findings underscore the need in future research to further refine fNIRS system design, notably montage geometry (e.g., asymmetry, number of channels) and the integration of short-distance channels.
Our study has several limitations that should be acknowledged. The primary limitation is the small sample size, which may affect classification robustness. Additionally, age distribution is not uniform across groups, which could introduce bias in the results. While no significant age difference was observed between the BP I and BP II groups, both were significantly older than HC.
Comparing our behavioral analysis results (no significant difference in accuracy) with those of Quan et al. [26] suggests that our images elicited effects typical of low motivational intensity stimuli. This intentional selection was made to avoid patient discomfort, which may have consequently led to a reduced response.
Finally, limitations in the fNIRS data preprocessing must be considered. Due to the absence of short-distance channels in our system, the PCA was used to minimize extracerebral activity [34]. However, hardware-based short-distance channels are more effective for this correction, as they account for local extracerebral activity [53, 54]. Another limitation concerns the fNIRS system we used, since it provided only six channels and therefore limited the spatial coverage of the recorded signals.
Overall, our results demonstrate that activity measures from cortical superficial brain regions provide robust diagnostic insights, yielding strong group classification performance. Furthermore, the combination of EEG and fNIRS enhances classification accuracy across all comparisons, with fNIRS providing valuable neurovascular insights that refine feature selection, particularly in distinguishing BP II from HC. Its integration compensates the limitations of restricted frontal EEG spatial coverage by capturing subtle neurovascular changes, thereby enhancing diagnostic accuracy and supporting the potential for a simplified and accessible diagnostic tool.
Further validation of these results will be conducted within a larger database, addressing cohort limitations and extending the analysis to vulnerable populations, such as relatives of BP. Although this group was excluded from the current study due to insufficient data, future research will investigate this population to identify early markers of emotional dysregulation. Additionally, newly acquired data will enable the examination of different BP mood states (depressive, euthymic, and manic) in relation to clinical scores. This will allow for a deeper understanding of how emotional regulation networks fluctuate across these states.
SUPPLEMENTARY DATA