acceptodds
Under review as a conference paper at ICLR 2027

BrainFlow: Depression Detection via Trajectory Deviation on Generative EEG Velocity Fields

Abstract

Electroencephalography (EEG)-based depression detection predominantly relies on discriminative representation learning that learns decision boundaries between healthy and depressed individuals. However, substantial inter-subject variability may entangle disease-related patterns with cohort-specific variations, limiting generalization and obscuring individual deviations from healthy EEG. To address these limitations, we propose BrainFlow, a generation-based framework that learns healthy EEG velocity fields from heterogeneous datasets to establish a transferable normative reference. Yet absolute mismatch to the learned healthy dynamics can remain unreliable, since even healthy subjects may exhibit considerable residuals due to individual variability, distribution shifts, and imperfect generative modeling. We therefore introduce Trajectory Deviation, which compares unseen subjects with locally matched healthy trajectories and corrects their transport residuals using corresponding healthy references. The resulting deviations are further calibrated into physiologically interpretable Flow-Energy representations for supervised depression detection. Extensive experiments demonstrate the effectiveness of BrainFlow under subject-independent evaluation settings.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, Depression Detection, Generative Velocity Fields, Trajectory Deviation.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zDFbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BrainFlow, a generative framework for EEG-based depression detection that reframes the task from direct HC/MDD classification to subject-specific deviation from healthy EEG dynamics. BrainFlow first learns healthy generative velocity fields from multi-source HC, then compares each unseen subject with locally matched healthy trajectories to compute relative trajectory deviations. These deviations are converted into physiologically traceable Flow-Energy features and used for final depression prediction. Experiments on five public EEG datasets show improved subject-independent performance and better interpretability of abnormal EEG patterns.

Critical strengths
  • Novel problem formulation. Instead of directly learning an HC/MDD decision boundary, BrainFlow models a healthy EEG generative reference and detects depression through subject-specific deviation from healthy dynamics, which provides a clear conceptual contribution.

  • Good consideration of inter-subject variability. The proposed trajectory deviation compares each subject with locally matched healthy trajectories rather than relying on absolute reconstruction or residual errors, which is well motivated for heterogeneous EEG data.

  • Relatively comprehensive experimental evaluation. The study integrates five public EEG datasets with 942 participants and adopts strict subject-level separation between healthy-reference learning, calibration, classifier training, and testing.

  • Strong interpretability. The Flow-Energy representation can be traced back to specific frequency bands, electrode node strengths, and regional functional connectivity, allowing subject-level physiological attribution of predictions.

Critical weaknesses
  • The framework still relies heavily on handcrafted EEG features. BrainFlow operates on 170 predefined wPLI connectivity and node-strength descriptors rather than learning representations directly from raw EEG, which may limit its ability to capture richer temporal and spatial information.

  • The pipeline is relatively complex. Healthy flow modeling, reference-bank construction, quality control, nearest-trajectory matching, HC calibration, Flow-Energy construction, and subsequent classification introduce many processing stages and design choices, which may affect reproducibility and practical deployment.

  • The performance improvement over strong EEG models is moderate. The proposed Flow Matching version achieves an F1 of 0.786, compared with 0.766 for ST-EEGFormer and 0.751 for EEGPT, so the empirical gain is meaningful but not particularly large given the added complexity.

  • The learned healthy reference may depend strongly on the diversity of available HC data. Although multiple datasets are used, the robustness of the normative velocity field to substantially different populations, acquisition systems, or clinical conditions is not yet fully established.

Questions for authors
  1. BrainFlow relies on handcrafted wPLI connectivity and node-strength features. Could the authors discuss whether the proposed healthy trajectory modeling can be extended to raw EEG or learned representations?

  2. The framework contains several design choices, including trajectory matching, HCQC-A filtering, probe times, and Q25 aggregation. How sensitive is the model to these choices, and which components are most critical in practice?

  3. The proposed method achieves only a moderate improvement over strong EEG baselines. Could the authors further justify whether the added complexity is warranted by the performance and interpretability gains?

  4. Since the healthy velocity field is learned from multi-source HC datasets, how robust is the learned normative reference when applied to populations or acquisition settings that differ substantially from those seen during training?

Reviewer_XPajOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this study, they extracted wPLI-based connectivity features using BrainFlow and learned the flow of the feature distribution using only the healthy group. For a new subject, they identified the most similar trajectory within the healthy reference set and calculated the residual difference between the subject’s trajectory—constructed from the same Gaussian starting point—and the corresponding healthy reference trajectory. After standardizing these residuals using statistics derived from the healthy group, they performed final classification into HC or MDD categories using logistic regression.

Critical strengths

The proposed method called BrainFlow involves the model learning not the temporal changes of raw EEG waveforms, but the process of generating a distribution of summarized EEG features for each participant. Since residuals can be large even in the normal group due to individual differences and model errors, it appears appropriate to approach the issue by comparing it with local health standards rather than judging abnormalities based solely on the size of the subject's residual.

Critical weaknesses

The flow mismatching method is a technique used in previous studies, and it appears that similar research has also been conducted in recent studies, such as "Relative Mismatch: Local-Reference Calibration of Feature-Space Flows for Anomalous Sound Detection." It is necessary to clearly articulate the differences between these approaches. Also, statistical analysis methods should be conducted for performance comparison.

Questions for authors
  • Even though various types of inputs can be derived from EEG data such as raw signals or spectrograms, the authors should sufficiently explain the rationale for choosing the "weighted phase-lag index" (wPLI). Please provide a clear justification for this choice.

  • The results in this paper are based on a single model architecture run with five different random seeds with no statistical significance analysis. If the authors intends to emphasize "repeated evaluation across different subject splits" or "transferability," a "leave-one-dataset-out" cross-validation approach. should be conducted.

  • The test set consists of 156 individuals with MDD and 22 HC. Given this class imbalance, high recall or F1-score for the MDD group is insufficient to determine how well the model distinguishes between the two groups. Please provide additional metrics, such as specificity, balanced accuracy, AUROC, and AUPRC. Also, please include a confusion matrix that would be beneficial for a comprehensive performance evaluation.

  • The 68.6% accuracy reported in the paper is an unweighted average of the accuracies obtained for each individual dataset. As results, the authors provided a direct comparison such as claiming that "predicting 'MDD' for every sample would yield 87.6% accuracy, which outperforms the proposed model". This would be a flawed assessment. Instead, the authors should present a simple baseline model evaluated using the same dataset-macro averaging method.

  • The performance evaluation regarding the model’s storage requirements, computational costs, and inference speed (e.g., FLOPS, total number of parameters, and inference latency) should also be investigated.

  • A recently published study, "Relative Mismatch: Local-Reference Calibration of Feature-Space Flows for Anomalous Sound Detection," also addresses mismatches caused by domain shifts by calibrating them against nearby "normal" references. Since this work appeared around the same time as your manuscript, it is not a matter of citating the paper. The authors should clarify the differences between the proposed method and this approach.

  • The proposed method can be interpreted as reweighting and combining feature differences between a subject and a normative reference, leveraging the local structure of the learned normative distribution such as an interpretation derived directly from the mathematical formulations in the paper. The comparison of "matched endpoint distances" in Table 2 serves as a good starting point for addressing this. In order to clarify whether the performance gains stem from the distance transformation, the learning of nonlinear structures, or the aggregation of data across multiple time points, two additional comparisons as shown below would be helpful. (1) nonlinear flow model versus an affine/Gaussian reference model (2) the use of nine probe time points versus a single probe time point.

  • Although the paper states that the Oz channel is not used, Figure 4 appears to include it. It would be beneficial to clarify this.

  • The information regarding baseline models is insufficient to determine the key parameters, making replication difficult. Details regarding the model architecture (e.g., the number of hidden layers) should be provided.

  • The current manuscript lacks an analysis of demographic distributions and covariates across different roles and diagnostic groups. These information is difficult to discern whether high "Flow-Energy" values ​​reflect differences associated with depression or, alternatively, demographic or technical mismatches between the subject and the normative reference. This issue should be discussed in the paper.

Reviewer_iCnKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper identifies two challenges in discriminative EEG-based depression detection. First, inter-subject variability is substantial. Second, conventional discriminative prediction only indicates which side of the decision boundary a subject falls on; it does not characterize how the subject differs from healthy controls (HC). To address both challenges, the authors propose BrainFlow, which learns a healthy reference through generative velocity fields and characterizes each subject based on deviations from that reference.

BrainFlow consists of two stages. Before these stages, EEG recordings are preprocessed and converted into handcrafted features: wPLI-based regional connectivity and channel node strength across five frequency bands, organized into ten feature groups. In Stage 1, a flow-matching velocity field is trained separately for each feature group using only HC subjects and is then used to generate a bank of 4,096 healthy reference trajectories. In Stage 2, each subject's features are matched to the K=8K = 8 nearest healthy endpoints in the reference bank. The subject and each matched reference are compared along paths originating from a shared Gaussian starting point under the learned velocity field. The resulting relative residuals are calibrated against HC statistics to produce "Flow-Energy" features, which are then used by an L2-regularized logistic regression classifier to predict MDD.

The experiments use five public EEG datasets (Dortmund, LEMON, MODMA-128, Mumtaz, and TDBRAIN), comprising 942 participants, with data partitioned at the subject level. Of these, 472 HC subjects are used to train the healthy velocity fields. A disjoint pool of 157 HC and 313 MDD subjects is used for calibration, classifier training, and testing, although the size of the test set is unclear. BrainFlow is compared against four families of models: conventional deep learning, flow-based, one-class, and EEG foundation models, although the training procedures for these baselines are not clearly described. In Table 1, BrainFlow with flow matching achieves the highest accuracy (0.686 ± 0.026) and F1 score (0.786 ± 0.019), as well as the second-highest recall (0.844, compared with 0.848 for the Transformer). However, its precision (0.753) is lower than that of seven baselines, including the one-class methods (0.88–0.91) and EEG foundation models (0.76–0.84). The flow-matching variant outperforms the diffusion variant across all four metrics (F1: 0.786 vs. 0.755; accuracy: 0.686 vs. 0.643).

Critical strengths

Well-motivated problem framing. The authors identify two major limitations of the current discriminative paradigm in EEG-based depression detection. First, EEG exhibits substantial inter-subject variability, which can cause distribution shifts even among subjects within the same dataset. Generalizing to unseen subjects therefore resembles a cross-domain generalization problem, and subject- or cohort-specific variations can become entangled with disease-related differences in the learned decision boundary. Second, a discriminative classifier only indicates which side of the HC/MDD decision boundary a subject falls on; it does not characterize how that subject's EEG differs from that of healthy controls. Reframing detection in terms of deviations from a transferable healthy reference directly addresses both limitations. This is an important strength of the paper because both challenges are widely recognized in the field and remain largely unaddressed by architectural improvements to discriminative models.

A novel approach to EEG-based depression detection. The proposed method introduces a new approach to EEG-based depression detection. Prior generative EEG research has primarily used GANs, VAEs, or diffusion models for data augmentation or representation learning. In contrast, BrainFlow uses a flow-matching velocity field trained exclusively on healthy controls to derive diagnostic evidence. The method combines this field with a bank of 4,096 healthy reference trajectories and a deviation-inference step. This step adjusts a subject's transport residual using its K=8K = 8 nearest healthy references, suppressing residual variation that is also present in healthy subjects. Because each Flow-Energy feature can be traced back to a specific EEG descriptor and the final classifier is linear, the approach also provides feature-level attributions by channel, region, and frequency band (Figs. 4–6). This is a meaningful contribution because the deviation-based representation, rather than the classifier itself, is the main innovation of the paper and could potentially be applied beyond depression detection.

Critical weaknesses

The use of velocity fields is neither justified nor validated. The paper does not explain why a generative velocity field is necessary or how it relates to depression-related EEG changes. The healthy velocity fields are trained on only 472 HC feature vectors, and the paper provides no evaluation of how well they represent the healthy distribution. The only related check, HCQC-A in Appendix A.5, involves per-feature marginal screening against held-out HC data. However, this procedure is used to discard poorly fitted features rather than being reported as an evaluation of the learned healthy distribution.

The experimental setup is unclear, and the reported gains are not convincingly demonstrated against a trivial classifier.

The test set is not adequately described in the main text. Section 4.1 only states that a disjoint pool of 157 HC and 313 MDD subjects is used for calibration, classifier training, and testing. It does not specify the test-set size or class balance across datasets. Only Appendix C.3.1 reveals that the test set contains 22 HC and 156 MDD subjects. This substantial class imbalance makes MDD-class F1 and accuracy potentially misleading. If the metrics were pooled over the reported test set (22 HC and 156 MDD), a trivial classifier that always predicts MDD would achieve an F1 score of 0.934 and an accuracy of 0.876. Thus, BrainFlow's reported performance offers little evidence of improvement over a classifier with no discriminative ability and may even be worse under this evaluation setting. No AUROC, balanced accuracy, specificity, or majority-class baseline is reported.

The training procedures for the baselines are insufficiently specified. The paper only states that all methods use the same non-test participants. It does not clarify which subjects and labels are used to train each baseline, how flow-based scores are converted into class predictions, or how the foundation models are fine-tuned. Several baselines (e.g., GNN, with an accuracy of 0.562) perform below the majority-class accuracy level, raising concerns that they may be insufficiently tuned.

Questions for authors
  1. Please report the number of HC and MDD subjects in the test set for each dataset, along with AUROC, balanced accuracy, and the performance of a majority-class (“always MDD”) baseline on the same test set. With 22 HC and 156 MDD subjects (Appendix C.3.1), a classifier that always predicts MDD would match or exceed the reported F1 score and accuracy. These additional metrics would help establish whether BrainFlow learns meaningful discriminative patterns beyond the class prior.

  2. Which subjects and labels were used to train each baseline? Most baselines would typically be trained using both classes, so were they trained on the same data splits as BrainFlow? How were the scores from the one-class and flow-based methods converted into MDD/HC predictions? Several baselines perform below the majority-class accuracy level (e.g., GNN, with an accuracy of 0.562). Without these details, it is difficult to assess whether the comparisons in Table 1 are fair.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.