PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Chanwoo Park, Chanwoo Kim

arXiv:2608.17231v1cs.LGcs.AI

TL;DR

Affordable, scalable dementia screening is needed because conventional imaging is costly and difficult to access. Delta2Gamma learns EEG representations through band-specific contrastive learning with adaptive temperatures, achieving 92.37% accuracy under strict leave-one-subject-out evaluation.

  • Problem

    Costly, specialized imaging limits accessible early dementia screening, motivating scalable alternatives such as EEG-based assessment.

  • Method

    Delta2Gamma decomposes EEG into five frequency bands and learns band-specific representations with independent contrastive heads and adaptively predicted temperatures.

  • Results

    92.37% accuracy under strict LOSO cross-validation surpasses recent dedicated Alzheimer’s disease-versus-control methods.

  • Takeaways & Limitations

    The framework delivers state-of-the-art Alzheimer’s disease-versus-control classification despite limited labels and is presented as extensible to other biosignal-analysis domains.

Abstract

from arXiv · show

Low-cost, scalable screening for dementia remains an open problem. Imaging-based diagnosis is costly and hard to deploy widely. Electroencephalography (EEG) is portable and inexpensive, but its recordings are noisy, vary widely across subjects, and carry few clinical labels. We tackle this with Delta2Gamma, a self-supervised framework that learns EEG representations from unlabeled data by contrasting augmented views of each signal. Rather than treat EEG as a single stream, Delta2Gamma decomposes every recording into the five canonical neural rhythms (delta, theta, alpha, beta, gamma). Each band gets its own encoder and projection head. Each also gets a temperature that is predicted adaptively during contrastive training, so bands with different signal statistics are balanced automatically. On the ADFTD cohort under a strict leave-one-subject-out protocol, Delta2Gamma separates Alzheimer's disease from cognitively normal controls with 92.4\% accuracy. This exceeds both supervised backbones and recent dedicated EEG methods.

I. INTRODUCTION

The introduction motivates scalable dementia screening by contrasting the limitations of costly, inaccessible imaging with EEG’s portable, direct measurement of brain function. It then presents Delta2Gamma, which preserves Alzheimer’s-related frequency-band information through independent processing of five canonical EEG rhythms.

  • Motivation: High-cost MRI and PET are difficult to deploy for early detection because they require expensive, stationary equipment and specialized personnel.These limitations reduce their accessibility for widespread screening.
  • Motivation: EEG offers a scalable, non-invasive alternative that directly and quantitatively measures declines in brain function through neuronal electrical signals.This direct measurement provides neurophysiological insight into cognitive decline.
  • Neurophysiological basis: Alzheimer’s disease is associated with increased delta and theta power and reduced alpha, beta, and gamma power, reflecting overall slowing of brain oscillations.The gamma reduction is linked to higher cognitive functions such as short-term memory.
  • Proposed approach: Delta2Gamma decomposes EEG into delta, theta, alpha, beta, and gamma bands, processing each with an independent CNN encoder and projection head.This frequency-band-specific design aims to preserve band-specific neural information and support dementia classification under strict subject-independent evaluation.

II. PROPOSED METHOD · A. Self-Supervised Multi-Band Architecture

Delta2Gamma uses a two-stage pipeline that learns robust EEG representations through self-supervised contrastive pre-training and evaluates them with a classifier. Its EEG-specific design processes five canonical frequency bands in parallel, using independent band-wise projection heads and channel-preserving temporal feature extraction.

  • II. PROPOSED METHOD: The framework has two stages: self-supervised pre-training on unlabeled EEG followed by linear evaluation with a classifier.Pre-training learns general signal characteristics, while linear evaluation uses the frozen pre-trained encoder for AD versus cognitively normal classification.
  • II. PROPOSED METHOD: Data augmentation creates transformed EEG views whose representations are trained to remain robust during contrastive pre-training.The method learns general signal characteristics from unlabeled recordings while making representations robust to transformations.
  • A. Self-Supervised Multi-Band Architecture: The method adapts SimCLR by pulling augmented views of the same sample together and pushing different instances apart.This contrastive objective is tailored to EEG and supports representation learning when unlabeled EEG is abundant but clinical labels are scarce.
  • A. Self-Supervised Multi-Band Architecture: Each EEG recording is decomposed into five canonical bands—δ, θ, α, β, and γ—and processed simultaneously as parallel views.Bandpass filters produce five views with shape [5, C, L] from a raw multi-channel signal of shape [C, L].
  • A. Self-Supervised Multi-Band Architecture: Each frequency band has an independent projection head for fine-grained feature learning.The independent heads allow the model to process the five frequency bands simultaneously while learning band-specific features.
  • A. Self-Supervised Multi-Band Architecture: A raw EEG input contains C=19 channels sampled at 500 Hz, with each recording represented as [C, L].Here, L equals seconds × 500 Hz sampling rate.
  • A. Self-Supervised Multi-Band Architecture: Parallel 1-D depthwise convolutions use kernel size 7, padding 3, and groups = C before batch normalization and ReLU.Processing each channel independently preserves its temporal patterns during band extraction.

B. Data Augmentation

Delta2Gamma creates two semantically identical but morphologically different views of each EEG signal through noise, scaling, masking, and occasional channel dropout. Contrasting weak and strong views treats same-signal pairs as positives and different-signal pairs as negatives, promoting robust, diagnosis-relevant representations.

  • Augmentation design: Each EEG signal yields two views using Gaussian noise, amplitude scaling, time- and frequency-domain masking, and occasional channel dropout.The augmentations use Gaussian noise with standard deviation 0.03, scaling in [0.8, 1.2], 10% masking, and a 10% chance of dropping 10% of channels.
  • Contrastive objective: The method contrasts a weak and a strong view to suppress nuisance variation while preserving band-specific structure relevant to diagnosis.This contrastive setup encourages noise- and transformation-robust representations.
  • Contrastive objective: Views of the same signal form positive pairs, whereas views from different signals form negatives in an instance-discrimination task.The pairing strategy drives the model toward robust representations from unlabeled EEG data.

C. Training Objective

The training objective aggregates per-band adaptive NT-Xent losses with temperature regularization. Dynamically predicted band-specific temperatures reflect distribution and learning difficulty, stabilizing contrastive training across EEG bands.

  • Training Objective: The pre-training loss aggregates adaptive NT-Xent losses across bands and includes temperature regularization.This objective applies contrastive learning separately to each band while regularizing the temperature parameters.
  • Training Objective: Each band learns separate positive and negative temperatures based on similarity to positive and negative pairs.The temperatures are learnable, and the negative-temperature index is selected by maximum negative-pair similarity.
  • Training Objective: Dynamic band-specific temperatures reflect distribution and learning difficulty, unlike fixed-temperature SimCLR.Matching contrastive-signal strength to each band’s statistics addresses the uniformity–tolerance trade-off and stabilizes training across differing power profiles.

D. Downstream Task

The downstream classifier uses a three-layer MLP attached to the pre-trained encoder, evaluated with either a frozen encoder or end-to-end fine-tuning.

  • D. Downstream Task: A three-layer MLP is attached to the pre-trained encoder for classification.Its hidden layers have sizes 512 and 256, followed by an output layer sized to the number of classes.
  • D. Downstream Task: The hidden layers use ReLU activations, batch normalization, and dropout rates of 0.3 and 0.2.These components are specified for the two hidden layers.
  • D. Downstream Task: The study compares linear evaluation with fine-tuning all parameters.Linear evaluation freezes the encoder, whereas fine-tuning updates the entire model.

III. EXPERIMENTAL SETUP

The experimental setup uses distinct optimization protocols for self-supervised pre-training and frozen-encoder linear evaluation. Both stages run for up to 100 epochs with early stopping on an NVIDIA RTX 4090 GPU.

  • Pre-training: Pre-training uses AdamW with batch size 64, learning rate 1 × 10−4, weight decay 1 × 10−5, cosine annealing with warm restarts, and adaptive NT-Xent.The adaptive NT-Xent temperature ranges from 0.05 to 0.5 with β=0.01.
  • Linear evaluation: Linear evaluation uses LOSO cross-validation with frozen encoder weights, AdamW, batch size 32, learning rate 1 × 10−4, and cross-entropy loss.The encoder weights remain frozen during evaluation.
  • Training schedule: Both stages run up to 100 epochs with early stopping using patience 10 on an NVIDIA RTX 4090 GPU.Early stopping is applied in both pre-training and linear evaluation.

A. Dataset

The study uses the publicly available ADFTD dataset of resting-state, eyes-closed EEG recordings from 88 participants. Classification focuses on Alzheimer's disease versus cognitively normal controls, using recordings acquired with 19 scalp electrodes at 500 Hz.

  • Participant cohort: The ADFTD dataset includes 88 participants: 36 with Alzheimer's disease, 23 with frontotemporal dementia, and 29 cognitively normal controls.Mean MMSE scores were 17.75 for AD, 22.17 for FTD, and 30 for CN.
  • Classification task: Classification focuses on Alzheimer's disease versus cognitively normal controls because frontotemporal dementia is difficult to identify from EEG alone.The 23 FTD subjects were not used for classification.
  • EEG acquisition: EEG recordings were resting-state and eyes-closed, collected with 19 scalp electrodes under the 10–20 system at 500 Hz.Signals were recorded using a Nihon Kohden EEG-2100 system with impedance below 5 kΩ.

B. Preprocessing and Segmentation · C. Evaluation Protocol and Metrics · IV. RESULTS AND DISCUSSION

The study preprocesses eyes-closed EEG with channel averaging, bandpass filtering, artifact removal, and 30-second segmentation, then evaluates generalization using leakage-resistant LOSO cross-validation and multiple metrics.

  • B. Preprocessing and Segmentation: A 6th-order Butterworth bandpass filter retains EEG activity between 0.5–45 Hz relevant to distinguishing AD from CN.The filtering follows average referencing across all channels.
  • B. Preprocessing and Segmentation: Average referencing is applied across all channels because referencing strongly affects amplitude measurements.
  • B. Preprocessing and Segmentation: Independent component analysis within MNE-Python removes ocular and muscular artifacts from the EEG recordings.
  • B. Preprocessing and Segmentation: Eyes-closed EEG is segmented into 30-second epochs, matching the standard epoch length used in sleep research.
  • C. Evaluation Protocol and Metrics: LOSO cross-validation trains on N−1 subjects and tests on the held-out subject, repeating across all subjects.This design prevents cross-subject data leakage.
  • C. Evaluation Protocol and Metrics: The evaluation reports accuracy, precision, recall, weighted F1-score, and AUC under substantial inter-individual variability.The protocol is intended to provide a stringent, realistic estimate of generalization.

A. LOSO Performance · B. Multi-Band Feature Analysis · C. Ablation Study

Under strict leave-one-subject-out evaluation, Delta2Gamma substantially outperforms common EEG baselines and dedicated Alzheimer’s disease methods. Its band-specific encoders learn distinct activation strategies, while ablations show benefits from contrastive pre-training, multiple projection heads, and adaptive modeling.

  • A. LOSO Performance: The evaluation compares Delta2Gamma with major EEG benchmarks implemented using Braindecode and prior Alzheimer’s disease-versus-cognitively-normal studies.
  • A. LOSO Performance: 92.37% accuracy and 92.33% F1-score are attained by the adaptive five-band-head model under strict leave-one-subject-out cross-validation.The model surpasses recent dedicated Alzheimer’s disease-versus-cognitively-normal methods.
  • B. Multi-Band Feature Analysis: Band-specific encoders learn distinct activation strategies matched to signal characteristics across the δ, θ, α, β, and γ rhythms.The δ encoder is concentrated on few features, θ is more uniform, α has the highest and most consistent activations, and β and γ are intermediate but irregular.
  • C. Ablation Study: 92.37% accuracy is achieved by the full model, compared with 62.54% for a CNN trained from scratch without self-supervision.This corresponds to a 47.70% relative gain, supporting contrastive pre-training when labels are scarce.
  • C. Ablation Study: 89.65% accuracy is obtained with a single projection head, versus 91.14% for five band-specific heads.The adaptive model further improves 3.03% relative over the single-head baseline.

V. CONCLUSION

Delta2Gamma is presented as a multi-head SimCLR framework for contrastive EEG representation learning, using independent CNN encoders and adaptive temperatures across five frequency bands. Its aggregated per-band contrastive objective delivers superior representations and state-of-the-art AD-versus-CN classification under strict LOSO evaluation, while remaining deployable and extensible.

  • Delta2Gamma presents a multi-head SimCLR framework for contrastive EEG representation learning.
  • The model uses independent CNN encoders and adaptive temperatures for each of the five frequency bands.A per-band contrastive loss is computed and aggregated across bands.
  • Aggregated per-band contrastive learning achieves superior representation quality and state-of-the-art AD-versus-CN classification under strict LOSO evaluation with limited labels.
  • The approach is readily deployable and extensible to other biosignal-analysis domains.
Loading 2608.17231v1…