Neurocomputing

Background and Objective Epileptic seizure prediction is a critical challenge requiring the discrimination of subtle preictal physiological changes from interictal brain activity. While deep learning has shown promise in this domain, existing models often face limitations due to small EEG datasets, high computational costs for training from scratch, and a lack of patient-independent generalizability. In this paper, we present a novel framework for EEG-based seizure prediction that leverages pre-trained Vision Transformers (ViTs) through custom architectural modifications and optimized re-training strategies. Methods Our primary contributions include: [bullet]CVIT-ESP: A family of vision transformer architectures that replaces standard patch embedding layers with custom N-dimensional CNN stages to refine EEG representations. [bullet] ESPFormer: A lightweight, custom-designed transformer specifically engineered to mitigate overfitting on limited-scale EEG datasets. We identified optimal fine-tuning combinations for transformer blocks by devising a heuristic search-space reduction strategy, significantly reducing the training complexity. We validated our methods using the patient-independent MLSPred-Bench, involving 12 diverse benchmarks with varying seizure prediction horizons. Results Results demonstrate a clear progression in performance: while prior ResNet and vanilla Transformer models achieved an AUC-ROC of 69.0%, our CVIT-ESP architectures achieved the highest performance with a maximum average AUC of 76.4%. Conclusions These findings suggest that adapting pre-trained ViTs with domain-specific CNN front-ends and strategic fine-tuning offers a robust, generalizable, and resource-efficient path forward for clinical seizure prediction systems. Our code is available at: https://github.com/pcdslab/CVitEsp and https://github.com/pcdslab/ESPFormer

Determining the amount of information a neuron receives per unit of time is key to understanding brain connectivity and how neural networks encode and transmit information. In particular, estimating this information flow by distinguishing between excitatory and inhibitory synaptic contributions is critical to understanding neural network function, as maintaining the excitation-inhibition (E/I) balance regulates neuronal excitability and circuit stability, whereas its disruption can lead to a plethora of brain disorders, including neurodegenerative and psychiatric conditions. However, because synaptic conductances cannot be measured directly, inverse methods are required to infer them from the membrane potential --a readily measurable quantity. Although partial solutions have been proposed, accurately estimating these conductances remains a significant challenge due to the complexity and diversity of the inputs. This is particularly true in the spiking regime, where neurons actively fire. In this work, we introduce a novel computational strategy that combines two critical metrics extracted from the time course of the membrane potential recording: the amplitude of the spike and the interspike interval. By using these quantities, the proposed method enables the accurate separation of excitatory and inhibitory contributions, yielding highly favorable results in the spiking regime. Author summaryQuantifying the continuous stream of inputs a neuron receives is key to understanding brain connectivity. Inside the brain, individual cells must maintain a tight balance between excitation and inhibition (E/I) to process information correctly, as any disruption in this equilibrium can impair its functionality. However, directly measuring the underlying excitatory and inhibitory synaptic conductances is technically challenging, and existing mathematical tools often fail when neurons enter their active firing regime. In this work, we introduce a novel computational strategy designed to extract and separate these time-varying conductances directly from the neurons spiking activity. By dynamically tracking just two accessible metrics - the amplitude of the spikes and the time intervals between them - our algorithm estimates both conductance profiles with high precision. Furthermore, we demonstrate that this procedure is highly robust against realistic experimental noise and data variability, providing an accessible framework that does not require complex hardware or an unfeasible number of repetitive experimental trials. By tracking changes in the E/I ratio of the synaptic input, this method provides an efficient approach to detecting pathological imbalances and understanding how local connectivity shapes cellular functionality.

Interictal epileptiform discharges (IEDs) are diagnostically important EEG abnormalities observed between seizures. This study addresses a conditional spatial-classification task where every analyzed four-second epoch had already been reviewed and confirmed by experts as containing an IED, and the model assigned that epoch to one of five predefined scalp-distribution categories (generalized, frontal, temporal, occipital, or centro-parietal). The analysis therefore does not evaluate IED-versus-non-IED detection. After preprocessing, 2,514 IED-labelled epochs were analyzed using identical stratified epoch-level partitions, SMOTE based training, 26 handcrafted features per included channel, and multiple machine-learning classifiers. A staged channel ablation compared 19-channel scalp EEG, 21-channel EEG with ECG, and the complete 29-channel input containing scalp EEG, referential, ECG, and EMG channels. The best EEG-only result was obtained with linear discriminant analysis (88.89% test accuracy). CatBoost achieved 93.25% on EEG with ECG channel and 94.44% with the whole channel set. All eight directly comparable classifiers showed numerically higher test accuracy after ECG channel was added; for CatBoost, the increase was 6.35 percentage points. In the EEG with ECG channel, CatBoost model on ECG channel on right and left arm received respectively 15.79% and 15.12% of normalized global SHAP attribution, and beta-band power was the leading of all features (18.76%). These SHAP values indicate model-specific predictive contributions and do not establish physiological biomarkers, causal autonomic mechanisms, or clinical localization. The findings support a limited methodological conclusion which is ECG-derived features were associated with improved internal epoch-level categorization of expert-confirmed IED epochs. They do not establish IED detection, artifact rejection, independent EMG effects, or generalization to unseen patients.

A novel mathematical framework to define the threshold of action potentials in excitable cells is presented. Unlike previously applied methods that rely on approximations or bifurcations, the approach focuses on the geometry of membrane potential trajectories. The changes in concavity during the upstroke of an action potential can be directly obtained from a time series of voltages. The concavity criterion is then extended to models based on autonomous dynamical systems where the changes in concavity can be obtained analytically from a curve of inflection points in phase space. The inflection point manifold defines a region required for excitability: all the orbits that cross it contain action potentials, and all the trajectories that contain action potentials are in it. This analytical principle can then be used to define excitability in a dynamical system, and also a measure of excitability that enables quantification and comparisons of excitability across dynamical system. The measure provides a way to compare the excitabilities of systems that model neurons with different electrophysiological phenotypes and consider different stimulus conditions. The traditionally vague physiological concept of electrical excitability is transformed into a rigorous analytical description by considering the time-dependent curvature of the membrane potential. The criterion is robust across smooth, single compartment models of electrical excitability and can be can be extended to single compartment models in higher dimensions, and multicompartment models as well.

This study presents improvements to the Hodgkin-Huxley (HH) models of ionic conductance and action potential generation. Sodium and potassium conductances are expressed by a single analytical formula describing the impulse response of a convolution of exponential distributions within a short-memory integration space. Treating transmembrane ion transit duration as a random variable, conductance profiles are interpreted as realizations of the probability density functions governing ionic movements. Applying the central limit theorem, the lognormal distribution emerges as the asymptotic profile of ionic conductances, constituting a fundamental primitive for such biosignals. A temporal state-transition paradigm describes the action potential waveform through four successive membrane potential transitions. Applied to electrophysiological recordings from lamprey reticulospinal neurons, this framework enables indirect estimation of key physiological quantities, including depolarization threshold, Nernst potentials, and net ion fluxes across the membrane. These advances open new perspectives for parameter estimation from experimental data and neuronal network simulation.

Seizure diaries are one of the most useful sources of information in the management of epilepsy, however patient engagement with them can be sporadic. Sustained participation with seizure diaries affects the completeness and reliability of self-reported data, so it is vital to be able to measure engagement. To facilitate this, we create a multidimensional engagement metric with which to characterize how patients interact with their seizure diary. We utilise data from the Helpilepsy, a seizure diary application, common features found in application engagement metrics in business settings, and well understood clinical features to do this. Clustering is then performed to isolate different user groups based on how engaged they are, and these groups are studied to understand what drives the differences in engagement. We found three groups emerge from the clustering: low, medium and highly engaged users. Investigating these groups further, we put together a ``profile" for highly-engaged users. We find that they tend to be older at the point of diagnosis, and have had epilepsy for longer than the other users. We also find they tend to have had more medications, have higher doses of common anti-seizure medications, and they have more medications typically given to those with refractory epilepsy. The implications for e-diary design are that more attention should be given to those newer to epilepsy in the onboarding phase. Also, engagement is not necessarily based on just the upload of seizures, with other features of an e-diary being important to be filled in.

Objective: Computational magnetoencephalography (MEG) interictal epileptiform discharge (IED) detectors have mainly used generalized MEG-only models, whereas clinical MEG interpretation routinely integrates simultaneous electroencephalography (EEG) and includes MEG-unique or MEG-dominant discharges. We developed a patient-specific EEG-MEG IED detector and evaluated event-wise prediction stability across models and the effect of adding EEG to MEG-based prediction. Methods: Seventeen patients undergoing clinical EEG-MEG evaluation for epilepsy were retrospectively analyzed. Clinically accepted dipole-review IEDs were treated as positive events, and nonannotated events were sampled as negatives. Logistic regression (LR), random forest (RF), and a lightweight three-dimensional ResNet were trained separately within each patient using EEG-only, MEG-only, and combined EEG-MEG (EMEG) inputs. Primary performance metrics were the area under the receiver operating characteristic curve (ROC-AUC) and average precision. Event-wise stability was assessed using rank disagreement, rank volatility, and class-aware distribution quotient analysis. Results: Aggregate discrimination was high across models and modalities. Median ROC-AUCs for EEG, MEG, and EMEG were 0.850, 0.890, and 0.880 for LR; 0.880, 0.860, and 0.910 for RF; and 0.920, 0.960, and 0.960 for ResNet. Despite comparable aggregate performance, event-wise analysis revealed model-dependent prediction behavior. ResNet showed significantly lower non-IED rank volatility than classical machine learning models and lower non-IED rank disagreement, particularly compared with RF. Adding EEG to MEG was associated with more favorable class-aware event-wise positioning in most events, while MEG-unique/dominant cases showed greater relative MEG contribution. Conclusions: Patient-specific EEG-MEG IED detection revealed clinically meaningful event-wise differences not captured by aggregate metrics. Simultaneous EEG complemented MEG-based detection, while MEG contribution remained prominent in MEG-dominant cases, supporting multimodal patient-specific IED event prioritization.

Surface electromyography (sEMG) offers a silent and wearable input modality, but its practical use is limited by variability across users and recording sessions. This study presents a compact CNN- Transformer model for decoding isolated handwritten digits from eight-channel sEMG signals. The model combines trainable signal preprocessing, convolutional feature extraction, and Transformerbased temporal modeling. It was evaluated on ten recordings from five participants using recordingseen classification, leave-one-recording-out (LORO) generalization, and few-shot adaptation. The model achieved a mean macro F1 score of 0.924 {+/-} 0.059 in the recording-seen setting and 0.619 {+/-} 0.252 under zero-shot LORO evaluation. Adaptation using two labeled trials per digit increased macro F1 to 0.828 {+/-} 0.112, while ten trials per digit achieved 0.925 {+/-} 0.053. The proposed architecture also outperformed classical and neural baselines in the controlled LORO benchmark. These results indicate that compact CNN-Transformer models, combined with lightweight target-recording calibration, provide a promising basis for adaptive sEMG-based input systems.

AO_SCPLOWBSTRACTC_SCPLOWElectroencephalography (EEG) is a promising tool for automated detection of mild cognitive impairment (MCI) and dementia, but comparisons across studies are limited by inconsistent datasets and evaluation protocols. This study benchmarks ten deep learning models across four resting-state EEG datasets and eight binary classification tasks using a unified preprocessing pipeline and five-fold subject-wise cross-validation. Each experiment was repeated ten times. SCCNet obtained the highest mean subject-level accuracy, sensitivity, and F1 score, while ShallowConvNet achieved the highest mean segment-level accuracy, specificity, and precision. Subject-level aggregation improved mean accuracy for all evaluated models, and performance varied substantially across datasets and diagnostic tasks. Higher computational cost did not consistently correspond to better classification performance, with several compact architectures remaining competitive with substantially larger models. The results provide a reproducible reference for comparing EEG-based dementia classification models under consistent subject-independent evaluation conditions.

EEG microstates are a distinct number of quasi-stable spatial distributions of brain activity. Microstate trajectories are strongly suspected to reflect the underlying neural mechanisms during information processing and are therefore also called the "building blocks" of human thought. In this study, we examined, if EEG microstate sequences can serve as potential triggers for a Brain-Computer Interface (BCI). To this end, a semi-supervised deep learning model architecture consisting of an LSTM-based autoencoder and a dense neural network was utilized to classify between left- and right-hand motor imagery EEG data, with the resulting classification output serving as the BCI trigger. On the one hand, this was done in a 2-step approach, in which the autoencoder and classifer have been trained separately. On the other hand, an end-to-end approach was employed, where training was performed by combining reconstruction and classification losses. Results show that the proposed model architecture was able to extract relevant features from microstate sequences and exploit them for within subjects and sessions classification. Applying transfer learning to session-to-session or across-subject transfer resulted in peak classification accuracies around 89%. We also investigated to what extent transfer learning has to be applied to reach considerable classification accuracies serving as the calibration time representative. We found that on average around 400s are needed for BCI calibration when emplyoing our approach to reach 80% classification accuracy. The present study signifies that the investigation of EEG microstate trajectories can be a promising approach for extracting BCI triggers, as it reduces the dimensionality of multi-channel recorded EEG signals to a distinct number of brain states over time. Deep learning methods, especially transfer learning, applied to EEG microstate trajectories seem promising regarding user-convenient and calibration-free BCIs in real-world applications.

Micro-electrode array (MEA) recordings are widely used to characterize functional connectivity in neural cultures and have gained traction for the analysis of human brain slices. However, the impact of graph construction methodology on the resulting network topology has not been systematically quantified. Here, we benchmark three methods - shared spiking activity, Pearson cross-correlation, and the spike time tiling coefficient (STTC) - across 37 recordings from human cortical slice cultures classified into low, moderate, and high activity groups. We show that method choice alone produces large topological differences (Cohens d = 0.86-1.14 for clustering coefficient, d > 1.0 for node count), while higher-order features such as modularity remain stable. Each method exhibits a distinct sensitivity profile: shared spiking detects activity-dependent changes primarily through network size, correlation uniquely captures clustering differences, and STTC combines strong biological sensitivity with negligible parameter dependence across lag windows (all d < 0.1). Within shared spiking, z-score normalization dominates all other parameter choices (d > 1.0 versus bin size effects of d < 0.23), functioning as an implicit analytical null model that fundamentally reshapes the edge set rather than merely rescaling weights. Inter-method edge overlap is low (Jaccard index 0.08-0.45) and activity dependent, demonstrating that these methods identify substantially different connections from identical data. Our results reveal that methodological choices including construction method, threshold, and normalization introduce hidden degrees of freedom with effect sizes comparable to the biological signals being measured. We provide practical recommendations for parameter selection, reporting, and cross-method validation in MEA-based network neuroscience. Author SummaryWhen we record electrical activity from brain tissue using grids of electrodes, we can ask how different sites influence one another and map the tissue as a network of connections. Thanks to novel culturing methods, this approach is increasingly used to study human brain slices. However, deciding what is "connected" is not well defined. Researchers use several different methods, and it has never been clear how much this choice shapes the network they end up describing. Here we compared three widely used methods on 37 recordings from human cortical slices spanning a range of activity levels. We found that the method alone can change the apparent structure of the network as much as real biological differences do. The methods frequently disagreed about which connections exist and some technical choices, including normalization techniques, had surprisingly large effects. Because these hidden choices can rival the biological signal, we provide this benchmarking work with practical recommendations for selecting, reporting, and cross-checking methods, so that network studies of brain tissue become more transparent, comparable, and reproducible.

This paper introduces a sensorimotor learning framework for a corticocerebellar network, grounded in the perspective of population dynamics. Using an optimal control theory approach, the cerebellum model enhances preparatory activity through premotor input, allowing the motor cortex to reach the desired initial conditions for movement more efficiently. Unlike traditional motor learning approaches that focus on acquiring new skills, this paradigm emphasizes automatization of already executable behaviors through repetition driven by intrinsic motivation. The proposed model is evaluated using a center-out reaching task, demonstrating that the role of the cerebellum is to shorten the preparatory period required for the successful execution of the movement. These findings suggest that corticocerebellar interactions play a crucial role in optimizing motor preparation, offering insight into the neural mechanisms underlying movement efficiency.

With 302 neurons and a rigorously characterized connectome, the nematode Caenorhabditis elegans represents a powerful model organism to study the fundamental roles of neuronal circuits in behavior. However, despite the breadth of research, many questions remain unanswered regarding how these organisms are able to successfully navigate their environment. Here, we present a biologically grounded dynamical circuit model for the investigation of sensory-guided behavior during C. elegans chemotaxis. Our mathematical model consists of the chemosensory neuron AWA, interneurons RIM and RIA, motor neurons, including SMDs and RMDs, and body wall muscles that provide proprioceptive feedback through stretch receptors. After optimization with an evolutionary algorithm, the model locomotes effectively toward a chemical attractant, realistically capturing nematode chemotactic behavior. Chemotaxis is ensured by sharp turns, which resemble the omega turns of living nematodes, as a key emergent property of the model. The sharp turning behavior is triggered by decreases in the concentration of the attractant. These result in reduced AWA activity, which in turn triggers disinhibition of RIM and subsequent changes in RIA oscillations. The ensuing coordinated changes in downstream motor neurons activity patterns produce sharp turns, which correct the nematodes path, so that the model worm heads toward the attractant, and remains at its proximity, after it reaches the gradient peak. The proposed framework, along with its emergent dynamics, provides new insights into the minimum requirements for C. elegans circuitry to display major features of its chemotactic behavior, including omega turns. In parallel, it generates experimentally testable hypotheses with respect to the participating neuronal elements.

Aging affects individuals at varying biological rates, prompting the development of the Brain Age Index (BAI) to quantify neurobiological health relative to chronological age and disease risk. While structural MRI has dominated brain age prediction, its high cost, immobility, and low temporal resolution restrict its clinical scalability and responsiveness to transient neurophysiological changes. Electroencephalography (EEG) offers a highly scalable, portable, and temporally precise alternative capable of capturing dynamic brain states. However, the transition of EEG-based models to clinical biomarkers is impeded by methodological limitations, including small or biased datasets, inconsistent preprocessing pipelines, and a distinct lack of interpretable machine learning approaches. To address these persistent challenges, this paper presents a comprehensive, open-source, end-to-end pipeline for large-scale EEG-based brain age modeling. Developed using the Temple University Hospital EEG Corpus (TUEG) the largest publicly available resting-state EEG dataset. The pipeline encompasses rigorous data engineering, reproducible preprocessing, and robust feature extraction. Following quality control and subject-level dataset partitioning to definitively prevent data leakage, exactly 41,181 recordings were successfully retained. Two independent feature sets were extracted: the Catch22 time-series characteristics and a comprehensive set of spectral, aperiodic, and non-linear dynamics from the CCS toolbox. The methodology evaluates seven regression models, optimized via Optuna for hyperparameter tuning, and integrates SHAP (SHapley Additive exPlanations) for transparent feature importance analysis. By making this infrastructure publicly available, this work lowers the barrier to entry for large-cohort studies, fostering reproducible development and clinical validation of dynamic brain age biomarkers.

Whole-brain functional connectivity, estimated from magnetoencephalography (MEG) data, provides a compact representation of long-range neuronal communication, making it suitable for predictive biomarker discovery. In this work, we propose a deep learning framework (FC-CNN) for predicting brain states from frequency-resolved functional connectivity estimates derived from resting-state MEG recordings. We systematically compare the performance of FC-CNN to that of conventional regression methods using amplitude and phase-based functional connectivity in the well-studied age-prediction task on the Cam-CAN cohort (n=576). We show that FC-CNN outperforms conventional approaches, and that, compared to phase synchronization, amplitude envelope correlation consistently leads to higher prediction performance. Moreover, we present quantitative evidence that the weights of a trained deep learning model can enable neurophysiological interpretation of the activity patterns that inform successful predictions. Our work demonstrates that the proposed approach successfully decodes brain states from MEG functional connectivity and is promising for discovery of predictive biomarkers for brain disorders.

Glioblastoma (GBM) is a highly aggressive brain tumor with an extremely poor 5-year survival rate of 6.9%, largely attributable to the lack of reliable biomarkers. While competing endogenous RNA (ceRNA) and copy number variation (CNV) analyses offer unique biomarker identification potential, current approaches neglect the integration of multiple regulatory mechanisms for biomarker detection. To address this limitation, we applied relational graph convolutional networks (RGCNs) to ceRNA and CNV knowledge graphs through a novel late fusion ensemble architecture. The proposed architecture outperformed baseline models and identified five novel biomarkers, including hsa-miR-196a and hsa-miR-224. Kaplan-Meier survival analysis and Cox regression indicated that the identified genes hold significant prognostic and diagnostic power. The early stratification of the Kaplan-Meier curves indicates the potential these genes hold for patient survival prediction. The results illustrate that a late fusion RGCN ensemble effectively captures complex gene interactions, overcoming limitations of existing models and providing a framework for biomarker discovery. The novel biomarkers serve as prospective targets for future GBM therapeutic development and candidates for non-invasive diagnostic assays.

Abstract EEG foundation models (EEG-FMs) are evaluated almost entirely on disease-discrimination accuracy. A clinical biomarker additionally requires measurement reliability, the stability of repeated measurements on the same individual, which regulatory biomarker frameworks treat as a prerequisite that discrimination does not imply. We asked whether frozen EEG-FM representations provide such stability, whether it is predictable from conventional model descriptors, and what information supports it. We measured test-retest reliability, disease discrimination, and representation distinctiveness for nine frozen representations: six EEG-oriented foundation models spanning masked, contrastive and predictive pretraining, handcrafted spectral features, and two general-purpose time-series models with no EEG exposure. All were evaluated under one preprocessing pipeline across two healthy retest cohorts, at roughly one month and two years, and three neurodegenerative cohorts. Reliability was measured in healthy adults only; the disease cohorts contribute cross-sectional discrimination. Reliability, quantified as the intraclass correlation coefficient (ICC), varied enormously (mean 0.08 to 0.76; coefficient of variation, CV, 53.0%) while disease discrimination, measured as area under the receiver operating characteristic curve (AUC), occupied a far narrower range across the same nine (AUC CV 5.4%), a roughly tenfold difference in relative dispersion, described rather than formally tested. We did not test formal AUC equivalence, so we describe discrimination as varying substantially less than reliability rather than as equivalent. The variation was not consistently explained by pretraining paradigm or domain among the models studied, and a model with no EEG exposure was among the most reliable tested. Alpha-band information contributed disproportionately to reliability, whereas theta-band information ranked first for Alzheimer disease and frontotemporal dementia discrimination, directionally consistent with established EEG evidence in both conditions. Only the reliability half of that contrast is individually significant. Subspace geometry and band ablation, two methodologically distinct analyses, both indicate that the two properties are partially, not fully, dissociable, and network architecture determines whether the dissociation is preserved, traded off, or jointly degraded across depth. Reliability showed no detectable association with discrimination, pretraining paradigm, or domain, and had to be measured directly. We recommend it become a standard evaluation axis for EEG-FM representations intended for longitudinal or biomarker use, and release a reproducible pipeline.

EegFun.jl is an open-source package for electroencephalography (EEG) analysis implemented in the Julia programming language. EegFun.jl provides a flexible framework for EEG research, covering data import from standard file formats, filtering and re-referencing, Independent Component Analysis (ICA) for artifact detection/correction, epoch extraction, and ERP averaging and visualisation. The Julia language provides the readability of a high-level scripting environment together with execution speeds comparable to compiled code. EegFun.jl combines interactive data visualization with high-performance execution, making large-scale analyses both efficient and easy. Here, we provide a brief overview and introductory tutorial of the core stages of the EEG analysis workflow to illustrate the packages capabilities. The package is freely available under the MIT license.

Cross-dataset generalization of EEG-based classification under weak, proxy-derived labels remains an open problem for altered-states research. We present a reproducible eight-dataset alignment pipeline that maps eight heterogeneous EEG corpora (712,832 windows; 697,906 with valid labels) to a common 14-channel EPOC+ montage with 63-dimensional spectral features, and we recover the real 1-9 arousal self-assessments for MAHNOB-HCI from session.xml metadata. As a benchmark, Random Forest classifiers are trained on seven source domains and evaluated on the held-out target under both zero-shot and 20%-participant few-shot calibration. The benchmark exposes two concrete methodological pitfalls rather than a performance result: (i) per-class recall shows every target collapsing to a single majority class, and (ii) a within-dataset upper-bound experiment (Table 3) shows that six of eight proxy label sets sit at or below three-class chance even when trained and tested on the same dataset, so the cross-dataset failure is a label-validity problem rather than a transfer-method problem. Across the eight targets (20 seeds, 8,000 evaluation windows per target), zero-shot accuracy averages 36.85% (95% CI 34.40-39.30) and calibrated 43.76% (41.77-45.75), but balanced accuracy stays at 33.01-35.62% (Cohen's kappa <= 0.068), i.e. at chance. The +6.91pp mean change is driven almost entirely by a single target, ds006437 (6.31% -> 60.60%): the median paired change across all 160 seed-pairs is 0.00pp, and after Holm-Bonferroni correction only ds006437 and ds004572 remain significant, the latter with a practically null effect (+0.39pp). Balanced accuracy stays between 33.01% and 35.62% and Cohen's kappa at 0.009 +/- 0.032, i.e. at or barely above three-class chance, while per-class recall shows six of eight targets collapsing to Deep (96.7-100% recall) and two to Light (68.5-99.1%). The collapse persists under SMOTE oversampling, under an EEGNet-v4 deep-learning baseline, and under CORAL and AdaBN feature alignment, which locates the bottleneck in proxy-label validity and class overlap in the feature space rather than in classifier capacity. We position this work as a preliminary methodological study: its contribution is a reproducible eight-dataset alignment pipeline, recovered MAHNOB-HCI arousal self-assessments, a quantitative estimate of split-leakage inflation, and a transparently reported negative result rather than a performance claim.

Motivation: Graph Neural Networks (GNNs) have gained increasing interest in the biomedical domain, as the integration of prior knowledge and deep neural networks has the potential to enhance insights into molecular processes and disease mechanisms. However, a comprehensive and systematic assessment of model architectures, data modalities, graph structures, and their performance for graph signal classification in the biomedical domain is yet to be performed. In order to close this gap, we conducted a benchmarking study on multiple GNNs on a Protein-Protein Interaction (PPI) network for Kidney Renal Clear Cell Carcinoma and Breast cancer subtype prediction, performing an in-depth investigation of architectures, incorporating skip connections and various data modalities. Results: While none of the GNNs outperforms the structure-agnostic Multi-Layer Perceptron baseline, all of them can handle bimodal data (gene methylation and expression) and offer the ability to gain explainability based on PPIs. We offer practical guidelines for applying GNNs to graph signal processing tasks specifically for cancer classification. Depending on the underlying dataset and PPI structure employed, models on different data modalities outperform others. Overall, we suggest using ChebNet, which tends to outperform the Graph Convolutional Network and the Graph Attention Network in cancer subtype prediction. We recommend using GNN architectures that employ a simple flattening readout layer, as they provide better classification performance and faster training time than those with global average pooling. Additionally, we tested residual connections, but they had only an insignificant impact on classification performance.