Introduction
Neurological disorders currently affect over one-third of the global population. While electrical brain signals like EEG and iEEG are vital for diagnosis, they are notoriously complex, noisy, and non-stationary. Deep learning (DL) offers a path toward automation, yet the research field has historically operated as a "Wild West," characterized by fragmented methodologies and inconsistent evaluation metrics. This systematic review of over 450 studies aims to unify these efforts and introduces BrainBenchmark, an open-source framework designed to standardize the evaluation of brain signal models.
The Diagnostic Workflow
The process of moving from raw brain activity to a clinical diagnosis involves several sophisticated stages. The authors categorize this workflow into three primary phases: signal collection/preprocessing, feature extraction, and final analysis.
- Signal Collection & Preprocessing: Raw recordings are cleaned through denoising, filtering, and normalization to remove artifacts.
- Feature Extraction: The field is shifting from manual statistical and spectral features toward automatic feature learning using DL architectures like CNNs, RNNs, and Transformers.
- Analysis & Diagnosis: Tasks are divided into sample-level classifications (e.g., identifying chronic conditions like Alzheimer's) and event-level detection (e.g., pinpointing the exact moment a seizure occurs).
The Generalization Gap and Partitioning Strategies
A major hurdle in clinical deployment is the "generalization gap." Many models perform exceptionally well when tested on the same individuals they were trained on (subject-specific), but fail when applied to new patients (cross-subject).
As shown in the partitioning strategies, Cross-Subject evaluation is the most clinically relevant but also the most challenging due to high data demand and computational burden. Bridging this gap is essential for creating "universal" models that can be used in real-world clinical settings without patient-specific fine-tuning.
The Rise of Self-Supervised Learning (SSL)
The most significant trend identified is the move toward Universal Pre-trained Models. By utilizing massive unlabeled datasets—some exceeding 40,000 hours—researchers are using SSL paradigms like Contrastive Learning, Predictive Coding, and Masked Autoencoding (MAE). These models learn the fundamental structure of brain activity, which can then be fine-tuned for specific disorders with minimal labeled data. This "tokenization" of brain signals, similar to how Large Language Models (LLMs) treat text, represents a paradigm shift in the field.
Key Results and Benchmarking
The review highlights varying levels of success across different conditions:
- Seizure Detection: While accuracy exceeds 95% on controlled datasets like Bonn, performance on complex clinical data like the Temple University Seizure Corpus (TUSZ) remains a challenge, with F1-scores often between 60-70% in cross-subject scenarios.
- Sleep Staging: SSL-based models are achieving 84-86% accuracy, rivaling traditional supervised methods.
- Psychiatric Disorders: Graph Convolutional Networks (GCNs) are particularly effective for Depression and Schizophrenia, often reaching >98% accuracy by modeling disrupted functional connectivity as a graph.
To address the lack of standardization, the authors' BrainBenchmark provides a unified platform featuring 8 representative models and 9 public datasets, setting a new bar for transparency and reproducibility.
Future Directions
Despite these advances, several challenges remain. Most DL models are "black boxes," lacking the interpretability required for clinical trust. Furthermore, there is a pressing need for lightweight models capable of running on wearable hardware for continuous monitoring. Finally, the field still awaits its "ImageNet moment"—a truly massive, multi-modal dataset that can catalyze the development of a foundation model for the human brain.
Summary updated