What your brain activity says about you: A review of neuropsychiatric disorders identified in resting-state and sleep EEG data

Electroencephalogram (EEG) monitoring devices and online data repositories hold large amounts of data from individuals participating in research and medical studies without direct reference to personal identifiers. This review explores what types of personal and health information have been detected and classified within task-free EEG data, investigating the privacy risks involved with openly available EEG data as machine learning methods improve re-identification capabilities.

The gist

Task-free EEG data, including resting state (REO/REC) and sleep stages, can be used by machine learning to classify various neuropsychiatric disorders such as Autism Spectrum Disorder, Parkinson’s disease, alcohol use disorder, Major Depressive Disorder (MDD), and insomnia with high classification accuracy.

Data Classification Findings

The review synthesized studies classifying medical disorders using task-free EEG data. Key findings regarding the types of data analyzed are:

1. Resting state EEG data has been used to classify a variety of disorders, including Autism Spectrum Disorder, Parkinson’s disease, alcohol use disorder, often requiring only 5 mins of data or less for high classification accuracy (>90%).

2. Sleep EEG data tends to hold classifiable information about sleep disorders such as sleep apnea, insomnia, and REM sleep disorder, but usually involves longer recordings or data from multiple sleep stages.

3. In terms of disorder classification, 24 disorders were primarily identified with resting state EEG data, with the most common being Major Depressive Disorder (MDD) or depressive disorders (5 studies).

4. Sleep data was primarily used to identify 12 disorders, with the most common being insomnia (10 studies).

Data Characteristics Comparison

The study compared the characteristics of resting state and sleep EEG data across various metrics:

- Sample Size and Channels:

Resting state studies tended to use more subjects, with an average of 229 participants per dataset, while sleep studies used about 40 participants per dataset. Resting state studies also tended to use more channels, with an average of 34.3 channels, compared to the sleep studies' average of 3.4 channels per dataset.

- Signal Time:

In the resting-state EEG studies, relatively low amounts of signal time were required to detect disorders (Figure 3). The minimum recording time used was 30 s, with a mean of 4.7 min. For sleep data, the average amount of recording time was 5.3 hours.

- Classification Accuracy:

Resting state had a mean of 85.7 % accuracy while sleep data had a mean of 90.3% accuracy.

Quality and Reproducibility Assessment

The quality analysis revealed that while many studies achieved high classification accuracies, many studies lacked critical information.

1. Quality assessment showed that 8 (31%) resting state studies and 6 (35%) of the sleep studies received a ‘somewhat’ or ‘no’ in question 7 about quality of the machine learning performed in the study.

2. Many papers were found to lack details required for replication, such as details about machine learning and recording time, or proper data splitting without subject overlap, which is important to prevent unintended information leakage.

Privacy Risks and Future Directions

The paper emphasizes that access to EEG can reveal sensitive personal health information, raising ethical concerns regarding re-identification.

1. Reidentification is possible using machine learning on both resting state and sleep EEG data; studies have shown the ability to reidentify individuals simply using their resting state (Zhang et al., 2018a; Zhang et al., 2020).

2. The risk is heightened if a direct personal identifier is never made accessible, as if just one data set reveals the identification directly or based on meta-information, other data sets could be de-anonymised by the mentioned methods.

3. To mitigate risks, efforts should focus on anonymization, such as the removal of all personalized and personal information such as an individual’s name and age from the research data, and developing tools for keeping privacy safe before machine learning advances allow for disorder classification in reidentified datasets.

4. Future methods may involve removing or disguising identifying information from the EEG data to anonymize it before publishing, or only release necessary non-identifying features of the data upon request.

Limitations

Despite high classification accuracies, limitations exist when applying these models to real-world scenarios:

1. There are differences between study datasets and a real world sample, and the base rate of many disorders is much lower than what is often observed in controlled studies.

2. Real-world data frequently includes individuals with comorbidity, or individuals with multiple disorders at the same time, whereas most studies listed only one disorder within each participant.

Based on the review of current research regarding task-free EEG data, here are specific improvements for AI systems, categorized by domain:

)1. Enhanced Diagnostic Classification Systems (Focus: Accuracy & Robustness)

The paper highlights that while classification accuracies are generally high (e.g., 90% for insomnia), many studies lack critical information regarding machine learning protocols (e.g., epoch length, training/testing split, subject overlap).

• Improvement: Develop Meta-Quality Check layers within the AI pipeline that automatically assess the metadata of incoming EEG datasets against established best practices (as outlined in Section 15).

• Improvement: Implement robust data splitting mechanisms that explicitly prevent subject overlap between training and testing sets to mitigate unintended information leakage, a known vulnerability identified in Section 16.

• Improved AI Capability: The system can achieve higher reliability when deployed on real-world, unseen datasets by ensuring the model generalizes correctly and is not biased by information leaked from the training set.

)2. Privacy-Preserving Re-identification Mitigation (Focus: Security & Anonymization)

The core concern is that task-free data, even without direct identifiers, can be used for re-identification via meta-information or sensitive features (Section 4).

• Improvement: Integrate Privacy Augmentation modules using techniques like Differential Privacy or Federated Learning specifically tailored for EEG signals. This would allow model training on distributed datasets without centralizing raw EEG data.

• Improvement: Utilize privacy-preserving neural network architectures, such as those employing Homomorphic Encryption (as mentioned in the references), to perform classification directly on encrypted EEG features.

-Improved AI Capability: The system can classify disorders or identify individuals from EEG data while mathematically guaranteeing that no identifiable personal traits (even subtle ones) are extracted or revealed by the model's parameters.

)3. Domain Adaptation for Real-World Deployment (Focus: Generalization & Bias Reduction)

The paper explicitly notes a limitation: real world data will often include individuals with comorbidity, or individuals with multiple disorders at the same time, as well as varying factors such as age, which were controlled in most of these studies (Section 18).

• Improvement: Develop Transfer Learning frameworks where models pre-trained on large, diverse datasets (like those from the Healthy Brain Network) are fine-tuned using small, specific clinical datasets.

• Improvement: Incorporate age and comorbidity as explicit, controllable input features into the model architecture (as opposed to relying solely on spectral features), allowing the AI to better handle real-world complexity.

-Improved AI Capability: The system moves beyond simple binary classification (e.g., Insomnia vs. No Insomnia) to a nuanced diagnostic tool that can predict complex multimorbidity patterns based on a patient's entire EEG profile, significantly improving clinical utility in diverse patient populations.

)4. Automated Feature Engineering & Selection (Focus: Efficiency & Reproducibility)

The review shows that different disorders benefit from different features (e.g., resting state vs. sleep data use more channels).

• Improvement: Implement an automated feature selection pipeline driven by domain knowledge, which dynamically selects the optimal set of EEG features (spectral power bands, connectivity measures, fractal dimensions) based on the specific disorder target and data modality (resting vs. sleep).

-Improved AI Capability: The system can automatically adapt its input requirements. For instance, when classifying Parkinson's disease from resting state data, it prioritizes motor cortex spectral features and higher channel counts (as suggested by Suuronen et al., 2023), while for insomnia classification from sleep data, it optimizes for temporal context learning (LSTM/CNN features).

Electroencephalogram monitoring devices and online data repositories hold large amounts of data from individuals participating in research and medical studies without direct reference to personal identifiers. This paper explores what types of personal and health information have been detected and classified within task-free EEG data. Additionally, we investigate key characteristics of the collected resting-state and sleep data, in order to determine the privacy risks involved with openly available EEG data. We used Google Scholar, Web of Science and searched relevant journals to find studies which classified or detected the presence of various disorders and personal information in resting state and sleep EEG. Only English full-text peer-reviewed journal articles or conference papers about classifying the presence of medical disorders between individuals were included. A quality analysis carried out by 3 reviewers determined general paper quality based on specified evaluation criteria. In resting state EEG, various disorders including Autism Spectrum Disorder, Parkinson's disease, and alcohol use disorder have been classified with high classification accuracy, often requiring only 5 mins of data or less. Sleep EEG tends to hold classifiable information about sleep disorders such as sleep apnea, insomnia, and REM sleep disorder, but usually involve longer recordings or data from multiple sleep stages. Many classification methods are still developing but even today, access to a person's EEG can reveal sensitive personal health information. With an increasing ability of machine learning methods to re-identify individuals from their EEG data, this review demonstrates the importance of anonymization, and the development of improved tools for keeping study participants and medical EEG users' privacy safe.