The efficacy and safety of antidepressants in bipolar depression have been debated for decades. Early observational studies suggested that antidepressants could effectively treat depressive episodes in bipolar patients, but randomized controlled trials have yielded mixed results [29]-[32]. The Systematic Treatment Enhancement Program for Bipolar Disorder (STEP-BD) found that adjunctive antidepressants did not outperform mood stabilizers alone for sustained recovery, though some subgroups may benefit [33]-[35].

The primary concern with antidepressant use is treatment-emergent affective switching (TEAS). Meta-analyses suggest that 15% - 30% of bipolar patients treated with antidepressants experience manic or hypomanic switches, with tricyclic antidepressants carrying higher risk than selective serotonin reuptake inhibitors (SSRIs) [36]-[39]. Risk factors for switching include previous antidepressant-induced mania, mixed features, rapid cycling course, and younger age at onset, but these clinical predictors have limited sensitivity and specificity [40]-[42].

Neuroimaging studies have identified several candidate biomarkers for bipolar disorder. Structural MRI studies consistently report reduced gray matter volume in the anterior cingulate cortex, ventral prefrontal cortex, and hippocampus [43]-[46]. Functional neuroimaging has revealed altered amygdala activation during emotional processing tasks, disrupted prefrontal-limbic connectivity during emotion regulation, and abnormal default mode network connectivity during rest [47]-[51].

Resting-state fMRI studies have particularly highlighted the importance of functional connectivity between the amygdala and prefrontal cortex. Bipolar patients often show increased amygdala reactivity to emotional stimuli coupled with reduced prefrontal regulatory responses, suggesting a neural substrate for emotional dysregulation that may also predict treatment response [52]-[55]. However, few studies have specifically examined fMRI predictors of antidepressant switching versus response.

EEG studies have identified several distinguishing features between bipolar and unipolar depression. Frontal alpha asymmetry reduced left frontal alpha power indicating relative left hemisphere hyperactivation has been reported in both conditions but may be more pronounced in bipolar depression [56]-[59]. Elevated theta power in frontal regions, increased beta activity, and reduced alpha coherence have also been described in bipolar patients [60]-[63].

Time-frequency analyses suggest that bipolar depression may be characterized by greater instability in neural oscillations, particularly in the theta and alpha bands, potentially reflecting the underlying mood instability that distinguishes bipolar from unipolar conditions [64]-[67]. These EEG signatures provide a rationale for using temporal deep learning models to classify depression subtypes.

Deep learning has increasingly been applied to psychiatric neuroimaging, with convolutional neural networks demonstrating superior performance over traditional machine learning for diagnostic classification [68]-[72]. In bipolar disorder specifically, CNNs have been used to classify patients versus controls based on structural MRI, with accuracies ranging from 70% - 85% [73]-[75].

For fMRI, 3D CNN architectures can capture spatial patterns of functional connectivity across the entire brain volume. Studies applying 3D CNNs to resting-state fMRI for depression classification have reported accuracies of 75% - 90%, though performance varies substantially across datasets and preprocessing pipelines [76]-[79]. Fewer studies have examined treatment prediction, though initial work suggests that deep learning can predict antidepressant response in unipolar depression with moderate accuracy [80]-[82].

For EEG, both 1D CNNs operating on raw time series and recurrent networks (LSTM/GRU) capturing temporal dependencies have shown promise for psychiatric classification. Studies comparing bipolar and unipolar depression using deep learning have achieved accuracies of 65% - 80%, with LSTM models often outperforming CNNs for capturing the temporal dynamics of neural activity [83]-[86].

Despite these advances, no prior study has integrated 3D fMRI CNN with EEG deep learning specifically for antidepressant decision-making in bipolar disorder. Our work addresses this gap by developing a multimodal framework that leverages the complementary strengths of spatial (fMRI) and temporal (EEG) neuroimaging.

To enable controlled evaluation and reproducible experimentation, we constructed physiologically grounded synthetic datasets for both fMRI and EEG modalities. Synthetic data generation allows systematic manipulation of neurobiological parameters while preserving realistic signal morphology, providing a stable testbed for model development and validation [87]-[91].

Label Definitions. Two distinct prediction tasks are implemented in this framework, each with separately defined labels. For the fMRI task: Antidepressant Responder (label = 0) refers to a synthetic patient profile representing a bipolar depression patient who, following antidepressant initiation as adjunct to a mood stabilizer, achieves ≥ 50% reduction in MADRS score within 12 weeks without any mood elevation event (YMRS increase ≥ 4 points from baseline). Treatment-Emergent Switcher (label = 1) refers to a profile representing a patient who experiences a transition to hypomania (YMRS 8 - 19) or mania (YMRS ≥ 20) within 12 weeks of antidepressant initiation, operationalized following the ISBD task force criteria. These two outcomes are mutually exclusive by design in the simulation: patients are categorized as switchers if any mood elevation event precedes or coincides with depression improvement. Non-responders who neither switch nor improve are not included in the current binary classification; future work will address this three-class problem. For the EEG task: Unipolar Depression (label = 0) refers to a DSM-5 Major Depressive Disorder profile without lifetime history of hypomania or mania. Bipolar Depression (label = 1) refers to a DSM-5 Bipolar I or II Disorder profile during a current depressive episode.

3.1.1. fMRI Data Generation

We generated synthetic resting-state fMRI volumes of size 64 × 64 × 64 voxels, representing standardized brain space. Each volume was constructed using a neurobiologically informed generative process that models key brain regions implicated in bipolar depression and antidepressant response:

Brain Regions Modelled:

Amygdala (left and right): Central to emotional processing and switching riskPrefrontal cortex (PFC): Executive control and emotion regulationAnterior cingulate cortex (ACC): Conflict monitoring and autonomic regulationHippocampus: Stress response and memory-emotion integration

Signal Generation Process: Each volume was initialized with a brain-shaped mask created using an ellipsoid approximation with added irregularity for realism. Regional activity was then superimposed as Gaussian activation spheres centered at anatomically appropriate coordinates:

For antidepressant responders (label = 0):

Moderate amygdala activity (intensity: 0.6 - 0.9)Balanced prefrontal activation (intensity: 0.7 - 1.0)Normal anterior cingulate response (intensity: 0.6 - 0.9)Stable hippocampal activity (intensity: 0.5 - 0.8)

For treatment-emergent switchers (label = 1):

Elevated amygdala hyperactivity (intensity: 1.1 - 1.5)Reduced prefrontal control (intensity: 0.3 - 0.6)Heightened cingulate activation (intensity: 0.8 - 1.2)Increased hippocampal reactivity (intensity: 0.9 - 1.3)

These parameters reflect the neurobiological hypothesis that switchers exhibit limbic hyperactivation coupled with deficient prefrontal regulatory control, while responders maintain more balanced prefrontal-limbic coupling [92]-[95].

Gaussian noise (σ = 0.15) was added to simulate scanner noise and physiological artifacts. Final volumes were z-score normalized to ensure consistent intensity distributions across samples.

3.1.2. EEG Data Generation

We generated synthetic 19-channel EEG data following the international 10 - 20 system, sampled at 256 Hz over 4-second epochs (1024 samples per channel). The generation process modelled characteristic oscillatory patterns distinguishing bipolar from unipolar depression:

Frequency Bands Modelled:

Delta (1 - 4 Hz): Slow-wave activityTheta (4 - 8 Hz): Frontal midline theta, linked to cognitive controlAlpha (8 - 13 Hz): Posterior dominant rhythm, inversely related to cortical activationBeta (13 - 30 Hz): Active cognitive processing

Channel-Specific Patterns:

For unipolar depression (label = 0):

Normal alpha power (25 - 35 μV²) with symmetric frontal distributionModerate theta activity (20 - 30 μV²)Reduced beta power (10 - 18 μV²)Balanced left-right frontal activation

For bipolar depression (label = 1):

Reduced alpha power (18 - 28 μV²) with frontal asymmetryElevated frontal theta (25 - 40 μV²)Increased beta activity (18 - 30 μV²)Reduced left frontal alpha (indicating relative left hyperactivation)

Frontal alpha asymmetry was implemented by reducing left frontal (Fp1, F3, F7) alpha power by 30% relative to right frontal channels, consistent with established findings in bipolar depression [96]-[99].

Signals were constructed by summing sinusoidal components for each frequency band with physiologically appropriate amplitudes, adding pink noise (1/f characteristic) for realistic background activity, and including 50 Hz line noise to simulate electrical interference. Final signals were z-score normalized.

3.1.3. Dataset Composition

fMRI dataset: 500 subjects (250 responders, 250 switchers)EEG dataset: 1000 subjects (500 unipolar, 500 bipolar)

Both datasets were split into training (70%), validation (10%), and test (20%) sets using stratified random sampling. To assess stability across data partitions, all three models (fMRI 3D CNN, EEG 1D CNN, EEG LSTM) were trained and evaluated across 10 independent random seeds (seeds 0 - 9), each producing a different stratified split. Results across seeds: fMRI 3D CNN mean accuracy 1.000 ± 0.000 (range 1.000 - 1.000); EEG 1D CNN mean accuracy 1.000 ± 0.000 (range 1.000 - 1.000); EEG LSTM mean accuracy 0.998 ± 0.004 (range 0.990 - 1.000). The near-zero variance across seeds confirms that perfect or near-perfect performance is not a consequence of a favourable fixed split but reflects the underlying non-overlapping class structure in the synthetic data, further reinforcing the interpretation that performance reflects generator design rather than model generalization capacity.

3.2.1. fMRI 3D CNN Architecture

We designed a 3D convolutional neural network to learn hierarchical spatial representations from volumetric fMRI data. The architecture progressively extracts features from local voxel neighborhoods to global brain patterns:

Convolutional Backbone:

Layer 1: Conv3D (1 → 32, kernel = 3, stride = 1) + BatchNorm + ReLU + MaxPool (2)Layer 2: Conv3D (32 → 64, kernel = 3, stride = 1) + BatchNorm + ReLU + MaxPool (2)Layer 3: Conv3D (64 → 128, kernel = 3, stride = 1) + BatchNorm + ReLU + MaxPool (2)Layer 4: Conv3D (128 → 256, kernel = 3, stride = 1) + BatchNorm + ReLU + AdaptiveAvgPool (4)

The progressive downsampling (64 → 32 → 16 → 8 → 4 voxels) captures multi-scale spatial patterns while the increasing channel depth (32 → 64 → 128 → 256) extracts increasingly abstract features.

Classification Head:

Flatten: 256 channels × 4³ voxels = 16,384 featuresDropout (0.5) → Linear (16,384 → 512) → ReLUDropout (0.3) → Linear (512 → 128) → ReLULinear (128 → 2) → Softmax

Total parameters: 9,618,562

3.2.2. EEG 1D CNN Architecture

The 1D CNN operates directly on raw EEG time series, learning temporal patterns through multi-scale convolutional filters:

Convolutional Backbone:

Layer 1: Conv1D (19 → 64, kernel = 51, padding = 25) + BatchNorm + ReLU + MaxPool (4) + Dropout(0.2)Layer 2: Conv1D (64 → 128, kernel = 25, padding = 12) + BatchNorm + ReLU + MaxPool (4) + Dropout (0.2)Layer 3: Conv1D (128 → 256, kernel = 9, padding = 4) + BatchNorm + ReLU + MaxPool (4) + Dropout (0.2)Layer 4: Conv1D (256 → 512, kernel = 3, padding = 1) + BatchNorm + ReLU + AdaptiveAvgPool (16)

The large initial kernel (51 samples ≈ 200 ms) captures slow oscillatory patterns, while progressively smaller kernels extract finer temporal features.

Classification Head:

Flatten: 512 channels × 16 samples = 8,192 featuresDropout (0.5) → Linear (8192 → 256) → ReLUDropout (0.3) → Linear (256 → 64) → ReLULinear (64 → 2) → Softmax

3.2.3. EEG LSTM Architecture

The LSTM model captures long-range temporal dependencies in EEG signals through recurrent processing:

Feature Extraction:

Conv1D (19 → 64, kernel = 25, padding = 12) + BatchNorm + ReLU + MaxPool (4)

This initial convolution reduces sequence length from 1024 to 256 while extracting local features.

Recurrent Processing:

2-layer bidirectional LSTM with hidden size 128Dropout (0.3) between layers

The bidirectional architecture processes the sequence both forward and backward, capturing past and future context for each timepoint.

Classification Head:

Concatenate final forward and backward hidden states (256 dimensions)Dropout (0.5) → Linear (256 → 128) → ReLUDropout (0.3) → Linear (128 → 2) → Softmax

Optimization:

Optimizer: Adam with weight decayLearning rate: 1 × 10⁻⁴ (fMRI), 1 × 10⁻³ (EEG)Weight decay: 1 × 10⁻⁵ (fMRI), 1 × 10⁻⁴ (EEG)Batch size: 8 (fMRI), 32 (EEG)Epochs: 30 (all models)

Loss Function: Cross-entropy loss for binary classification: L = −[y log(p) + (1 − y) log(1 − p)]

Regularization:

Dropout: 0.2 - 0.5 across layersBatch normalization after each convolutionEarly stopping based on validation loss (patience = 10 epochs)

Performance was quantified using:

Accuracy: (TP + TN)/(TP + TN + FP + FN)Precision: TP/(TP + FP)Recall (Sensitivity): TP/(TP + FN)F1-Score: 2 × (Precision × Recall)/(Precision + Recall)AUC-ROC: Area under the receiver operating characteristic curve

Given the clinical importance of identifying switchers (avoiding false negatives), we particularly emphasize recall for the switcher class.

Before evaluating predictive performance, we verified that the synthetic datasets exhibit physiologically meaningful structure and class-dependent variability.

Figure 1Figure 1 presents sample fMRI slices comparing antidepressant responders (top row) versus treatment-emergent switchers (bottom row) at four axial planes (Z = 32, 37, 42, 47). Visual inspection reveals distinct activation patterns: responders show more diffuse, moderate-intensity activation across prefrontal and limbic regions, while switchers exhibit focal, high-intensity activation particularly in subcortical structures. These patterns align with the neurobiological hypothesis of limbic hyperactivity in switch-prone patients.

Figure 1. Representative axial fMRI slices of antidepressant responders (top row) and treatment-emergent switchers (bottom row) at Z = 32, 37, 42, and 47.

Figure 2Figure 2 displays representative EEG traces from frontal (F3, F4), central (Cz), and occipital (O1) channels comparing unipolar (blue) versus bipolar (red) depression. Bipolar traces show visibly greater amplitude variability and reduced rhythmicity, particularly in frontal channels, consistent with the frontal alpha asymmetry and elevated beta activity programmed into the generation process.

Figure 2. Synthetic EEG time-series signals from frontal (F3, F4), central (Cz), and occipital (O1) channels comparing unipolar and bipolar depression.

The 3D CNN achieved perfect classification performance on the held-out test set. (See Table 1)

Table 1. fMRI 3D CNN results.

Figure 3Figure 3 shows the training dynamics over 30 epochs. Training loss decreased smoothly from initial values to near-zero, indicating effective optimization. Test accuracy reached 100% by epoch 5 and remained stable throughout training, with a transient drop to 80% at epoch 19 followed by rapid recovery. This stability suggests the model successfully learned robust discriminative features rather than memorizing training examples.

Figure 4Figure 4 presents the confusion matrix for the test set, showing perfect classification with 50/50 responders and 50/50 switchers correctly identified. No false positives or false negatives were observed.

Figure 3. Training and test loss curves and test accuracy over 30 epochs for the fMRI 3D CNN model.

Figure 4. Confusion matrix of the fMRI 3D CNN on the held-out test set.

Figure 5Figure 5 displays the ROC curve with AUC = 1.000, indicating perfect discrimination across all possible decision thresholds. The curve rises vertically to maximum true positive rate at zero false positive rate, reflecting the complete separability of the two classes in the learned feature space.

Figure 5. Receiver operating characteristic (ROC) curve for the fMRI 3D CNN classifier (AUC = 1.000).

Both EEG models demonstrated strong classification performance, with the LSTM achieving marginally superior results (See Table 2, Table 3).

Table 2. 1D CNN performance.

Table 3. LSTM performance.

Figure 6Figure 6 shows training curves for the 1D CNN. The model achieved rapid convergence, reaching 100% test accuracy by epoch 4. Training and validation loss curves track closely, indicating good generalization without overfitting.

Figure 6. Training and test loss curves and test accuracy for the EEG 1D CNN model.

Figure 7Figure 7 presents LSTM training dynamics. Similar to the CNN, the LSTM converged quickly to perfect test accuracy, with stable performance maintained throughout training. The recurrent architecture successfully captured the temporal dependencies distinguishing bipolar from unipolar depression.

Figure 7. Training and test loss curves and test accuracy for the EEG LSTM model.

Figure 8Figure 8 provides a comprehensive ROC comparison across all three models. All architectures achieved AUC = 1.000, demonstrating that both spatial (fMRI) and temporal (EEG) neural signatures are completely separable under the current synthetic data generation parameters.

The perfect performance across all models validates that:

1) The 3D CNN architecture can effectively learn volumetric brain patterns predictive of antidepressant response

2) Both CNN and LSTM architectures can capture EEG signatures distinguishing bipolar from unipolar depression

3) The synthetic data generation process creates physiologically meaningful, class-separable patterns

Figure 8. ROC curve comparison of all models: fMRI 3D CNN, EEG 1D CNN, and EEG LSTM.

The perfect classification performance (AUC = 1.000) across all models reflects the controlled nature of synthetic data generation, where class-specific patterns were explicitly programmed with minimal overlap. While these results demonstrate the feasibility of the proposed deep learning architectures, they should be interpreted as proof-of-concept rather than expected performance on real clinical data.

In real-world applications, several factors will degrade performance:

Individual variability: Real patients exhibit substantial heterogeneity in brain structure and functionScanner artifacts: Motion, physiological noise, and scanner drift introduce variabilityMedication effects: Psychotropic medications alter neural activity patternsComorbidities: Anxiety, substance use, and medical conditions affect neuroimaging signalsState-dependent effects: Mood state at scanning affects resting-state connectivity

Based on published literature, we anticipate real-world performance in the range of:

fMRI 3D CNN: 65% - 75% accuracy for treatment response predictionEEG models: 60% - 70% accuracy for bipolar/unipolar classification

These more modest accuracies would still represent clinically meaningful improvement over current decision-making, which relies primarily on clinical history with limited predictive validity.

If validated on real clinical data, this framework could support several clinical applications:

1)Antidepressant Risk Stratification: The fMRI model could identify patients at high risk for treatment-emergent switching before antidepressant initiation. For these patients, clinicians might prioritize mood stabilizer optimization, atypical antipsychotics, or psychosocial interventions over antidepressant trials.

2)Diagnostic Clarification: The EEG models could help distinguish bipolar from unipolar depression in patients with ambiguous clinical presentations. This distinction is critical given the divergent treatment implications: antidepressants are first-line for unipolar depression but potentially harmful in bipolar disorder.

3)Personalized Treatment Selection: By combining fMRI and EEG predictions, clinicians could develop personalized treatment algorithms that match patients to interventions based on neurobiological profiles rather than symptom history alone.

Integrated Clinical Decision Workflow. The fMRI and EEG models address complementary questions in the antidepressant decision pathway for a single patient. Step 1, Diagnostic clarification (EEG task): when a patient presents with depression of uncertain polarity (unipolar vs. bipolar), the EEG model is applied first. If the EEG model predicts bipolar depression with confidence ≥ 70%, the clinician is alerted to the possibility of a bipolar diagnosis and guided toward diagnostic confirmation. Step 2, Switching risk stratification (fMRI task), once a bipolar diagnosis is established or sufficiently likely, the fMRI model is applied to estimate individual switching risk. If the fMRI model predicts high switching risk (predicted probability for switcher class ≥ 0.65), the clinician is recommended against antidepressant initiation and guided toward mood stabilizer optimization or atypical antipsychotic alternatives. If switching risk is low (<0.35), cautious antidepressant initiation with close monitoring is supported. Intermediate risk (0.35 - 0.65) triggers a recommendation for enhanced monitoring and shared decision-making. This two-step workflow requires that both models are applied in sequence for the same patient, using resting-state fMRI (acquired at a single session prior to treatment) and a 4-second resting-state EEG epoch from the same session. Future work will validate this combined decision pathway on prospective clinical data.

Synthetic-Only Validation: All experiments used synthetically generated data. While designed to mimic realistic neurobiological patterns, synthetic data cannot capture the full complexity of clinical populations. External validation on real neuroimaging datasets (e.g., STEP-BD, UK Biobank) is essential before clinical consideration.

Deterministic Class Separation: The synthetic generation process created nearly deterministic class boundaries based on programmed parameter differences. Real neuroimaging data exhibits substantial overlap between clinical groups, and perfect separation is neither expected nor observed in published studies.

Simplified Neurobiological Model: Our generation process modelled only a subset of brain regions and frequency bands implicated in bipolar disorder. Real neural dynamics involve complex interactions across hundreds of regions and multiple oscillatory frequencies.

Absence of Confounding Variables: Real clinical data includes numerous confounding factors (age, sex, medication status, comorbidities) that were not modelled in the synthetic data. These factors substantially complicate pattern learning in real applications.

Real-World Validation: Priority should be given to validating these models on existing clinical datasets with treatment outcome data. The STEP-BD study, EMBARC trial, and similar multi-site studies provide opportunities for external validation.

Multimodal Integration: Future work should explore fusion of fMRI and EEG data within unified architectures, potentially using attention mechanisms to weight modalities based on their predictive utility for individual patients.

Longitudinal Prediction: Current models predict static outcomes. Future work should develop temporal models that track neuroimaging changes during treatment to predict emergent switching before clinical manifestation.

Interpretability Enhancement: The current framework does not include an implemented attention module or saliency procedure. The Abstract and Introduction references to “attention visualization” and “regional activation mapping” describe planned future work rather than completed analyses. Specifically, Gradient-weighted Class Activation Mapping (Grad-CAM) will be applied to the 3D CNN to generate voxel-level saliency maps identifying which brain regions most influenced each classification decision. For EEG, a temporal attention layer will be inserted between the convolutional backbone and the classification head, with attention weights visualized across the 1024-sample time axis to identify which temporal segments (and by Fourier analysis, which frequency bands) drove the classification. These interpretability enhancements are prerequisite for clinical credibility and will be implemented in the next study phase using real neuroimaging data.