Delta2Gamma: Band-Wise Adaptive Contrastive Learning of EEG for Alzheimer's Disease Detection

Interpretable self-supervised transformers for resting-state EEG analysis in Alzheimer's disease

Existing EEG-based methods have been constrained by limited availability of labeled data, hand-crafted features, poor spatio-temporal modeling, sub-optimal cross-hardware performance, and lack of interpretability due to being expensive and intrusive. To address these limitations, this study introduces innovative neural signal decoding techniques for cognitive state modeling in order to enhance the potential of AI-integrated models for early detection of Alzheimer's disease. This research introduces a self-supervised spatio-temporal transformer (STT-EEG) for early detection of Alzheimer's disease from resting-state EEG. This framework includes four key components: first, self-supervised pretraining on 111 healthy controls using masked auto-encoding and temporal order prediction to learn robust generalisable representations. Second, a novel spatial attention module (SAM) that explicitly captures both long-range temporal dependencies and channel interactions, reflecting the distributed network pathology of AD; third, cross-dataset transfer learning from 64-channel BioSemi to 19-channel Nihon Kohden systems, which showed strong hardware generalization; and finally, analyses of attention rollout and channel perturbation for clinically interpretable insights. The model was trained in a subject-wise 5-fold cross-validation fashion on the SRM dataset and fine-tuned on the OpenNeuro dataset (ds004504) consisting of 36 AD and 29 CN. On the same dataset, the accuracy of STT-EEG was 96.42% for AD vs. CN classification. The most significant improvement +7.08% was made with the help of self-supervised pretraining, followed by data augmentation +6.30% and the SAM +4.86%, as was confirmed in the ablation studies. For continuous prediction of MMSE scores, the Pearson correlation of the model was 0.872 and the mean absolute error (MAE) was 2.34 points. Regions of T3–T6, P3–Pz–P4 and O1–O2 were identified as areas of attention-based interpretability, which were consistent with the known neuropathology of AD that involved the temporoparietal lobe. STT-EEG strengths include the high interpretability of the framework, its spatio-temporal attention, and the fact that STT-EEG is a self-supervised learning method and can be used in an efficient and generalizable way.

Interpretable AI for neural signal decoding in dementia: an EEG ensemble approach to differential diagnosis

The differential diagnosis between Alzheimer’s disease (AD) and frontotemporal dementia (FTD) presents a significant clinical challenge due to overlapping early-stage symptom profiles. Conventional resting-state EEG provides limited sensitivity to the impaired neural plasticity and lateralized cortical degeneration frequently observed in FTD. We developed a domain-informed heterogeneous ensemble framework incorporating dynamic neural reactivity and hemispheric asymmetry metrics from 19-channel EEG recordings acquired from 88 participants (36 AD, 23 FTD, 29 cognitively normal controls) during resting-state and photic stimulation paradigms. A neural reactivity vector (V_diff) was derived to quantify state-dependent spectral transitions. The 1,014-dimensional feature space was reduced to 200 features via recursive feature elimination, prioritizing spectral power distributions, hemispheric asymmetry indices (HAI), and stimulation-induced reactivity parameters. A weighted ensemble of Extreme Gradient Boosting (XGBoost) and Random Forest classifiers was evaluated using a subject-aware 90/10 holdout split with internal five-fold cross-validation. The optimized model achieved a multi-class segment-level accuracy of 95.63% on an independently held-out test partition of 1,281 segments, with an internal five-fold cross-validation mean of 0.9846 ± 0.003. FTD-specific precision reached 0.9907. SHAP analysis identified beta-band hemispheric asymmetry and alpha-band reactivity as the principal contributors to class separation. These findings indicate that the integration of dynamic state-transition measures with structural asymmetry proxies enhances electrophysiological discrimination between dementia subtypes. The framework provides a computationally efficient and biologically interpretable alternative to deep learning–based methodologies for EEG-driven dementia classification.

EEGDecoder-x: an explainable deep learning framework for cross-subject EEG-based detection of Alzheimer's and Creutzfeldt–Jakob disease

Early detection of neurodegenerative diseases is critical. Distinguishing early-stage Creutzfeldt–Jakob disease (CJD) from “mimics” like Alzheimer's disease (AD) remains a major challenge; while EEG is valuable in advanced CJD, early-stage abnormalities are often non-specific and overlap with other rapidly progressive dementias. Deep learning offers promising EEG-based diagnostic solutions, but clinical adoption requires transparent decision-making, the interpretability of the features learned by deep learning models is equally important. In this context, careful model design and explainability are essential. In this paper, we propose a novel interpretable framework, EEGDecoder-x, for decoding EEG signals from subjects with Alzheimer's disease, Creutzfeldt–Jakob disease, and healthy controls, while providing insight into the model's learned characteristics. The EEGDecoder-x framework comprises two main components: a hybrid attention network for disease decoding (EEGDecoder-Net) and an explainability module (EEGDecoder-XAI). EEGDecoder-Net combines a convolutional neural network with a dual attention mechanism, followed by a classification layer, enabling efficient spatio-temporal feature extraction. EEGDecoder-XAI provides a comprehensive local and global explanations of the network's learning process for spatio-temporal dimensions. We validate the proposed framework using a Leave-One-Subject-Out evaluation paradigm, achieving 97.22% classification accuracy on a dataset of 36 subjects (12 with AD, 12 with CJD, and 12 healthy controls), and outperforming the baseline models, demonstrating both the effectiveness and interpretability of EEGDecoder-x.

TF-JointMAE: a self-supervised multi-representation EEG learning framework for Alzheimer’s disease spectrum classification

Prior-Informed Normative EEG Scoring for Alzheimer's Detection: Framework and Cross-Paradigm Boundary Analysis

EEG-based screening for Alzheimer's disease (AD) is attractive in principle but not deployed in practice, largely because end-to-end classifiers trained on small clinical cohorts generalize poorly. We propose a prior-weighted deviation scorer that fits only the healthy distribution and scores new recordings against directional priors drawn from the AD biomarker literature. Our choice of EEG paradigm was a crucial decision: in a controlled simulation with synthetic AD and synthetic depressionlike cohorts, the between-emotion variance of the deviation score separates simulated AD from simulated depression at AUC 1.00 while a resting-state analog (mean score only) sits at chance (AUC 0.49), since AD's spectral signature stays invariant across emotional conditions while affective pathologies do not. We therefore implement the approach on the SEED-VII emotion-task dataset $(n=20)$ using seven directional priors (no AD-labeled data in training). A deviation score is produced and algebraically decomposed into per-band and per-region contributions. As a pipeline-wiring check, simulated AD recordings constructed along the same prior directions are separable from healthy windows, while sign-flipped and undirected baselines score near chance - confirming that the scorer implements the literature priors as written. Within paradigm, a second healthy cohort (SEED-V, $n=16)$ is not distinguishable from SEED-VII at this sample size (Mann-Whitney $U, p=0.162)$. The headline result is across paradigm: zero-shot transfer to a resting-state clinical EEG dataset (ds004504, $n=88)$ returns AUC $=0.478$, with the three diagnostic groups clustering within 0.02 of each other, even though the raw band powers carry the expected AD signature (delta and theta elevation, alpha reduction). The framework holds within paradigm and fails across it; prior-weighted EEG scorers are paradigm-conditional.

Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease

Biological heterogeneity in Alzheimer's Disease (AD) poses a critical diagnostic challenge, particularly for traditional linear methods that fail to capture non-linear neural dynamics. To address this, we propose a diagnostic framework utilizing the Large Brain Model (LaBraM), pretrained on over 2,500 hours of EEG data. By integrating these high-dimensional latent embeddings with a non-linear Random Forest classifier, our approach effectively isolates robust disease markers. Under a rigorous subject-independent 5-fold cross-validation protocol, the method achieves an ROC-AUC of 89.36% +/- 3.49%, PR AUC of 81.45% +/- 4.43%, and Balanced Accuracy of 82.44% +/- 4.34% in distinguishing dementia patients from healthy controls. Notably, this performance uses only 8-second EEG segments, surpassing traditional spectral baselines, including band-power and parameterized oscillatory features (FOOOF). Post-hoc occlusion analysis confirms the model captures clinically validated biomarkers, specifically occipital-frontal Alpha and Theta rhythm degradation. Additional neurophysiological alignment analysis demonstrated that higher LaBraM-predicted dementia probability significantly correlated with worse cognitive performance, greater clinical severity, increased theta and alpha relative power, and higher aperiodic exponent. These findings demonstrate that deep latent representations extract clinically relevant signatures from noisy signals, enabling precise, rapid, and data-efficient diagnosis.