Can Artificial Intelligence Read Brain Waves and Recognize When Its Predictions Are Uncertain?

Researcher Rakesh Kumar Reddy AMBATI NeuroMCP-Agent combines brain-wave analysis, advanced machine learning, and responsible AI techniques to explore  a more  transparent approach to neurological and mental-health screening

Diagnosing conditions such as epilepsy, Parkinson’s disease and depression can be challenging because their symptoms often overlap. Rakesh Kumar Reddy Ambati, an Application System Analyst at Dartmouth Health, has developed NeuroMCP-Agent, an artificial intelligence system designed to analyze electroencephalogram (EEG) recordings and screen for seven conditions: Parkinson’s disease, epilepsy, schizophrenia, chronic stress, Alzheimer’s disease, autism spectrum disorder and major depression.

Beyond predicting conditions, the system aims to estimate confidence, explain the features influencing its decisions, and assess fairness, privacy, safety and robustness. These considerations are important in healthcare AI, where transparency and reliability matter alongside accuracy. However, the reported findings represent research results, not evidence of readiness for routine clinical use.

Why Brain Waves?

EEG records electrical activity in the brain through sensors placed on the scalp. Although non-invasive, EEG signals can be affected  by movement, eye blinks, muscle activity and recording conditions.

NeuroMCP-Agent processes these signals using filters and noise-removal techniques, divides recordings into four-second windows with 75% overlap, and extracts 47 measurements describing brain-wave frequencies, signal complexity and irregularity. The study also reports expanding training data fifteen-fold through generated variations to address differences  in available recordings. However,  augmented data do not represent new independent patients, making careful separation of training and testing data essential to prevent data leakage.

Combining Algorithms and Estimating Uncertainty

The system uses an “Ultra Stacking Ensemble” that combines 15 algorithms, selects 300 informative outputs and integrates them through a final model. It also uses Monte Carlo dropout to estimate uncertainty by generating repeated predictions with small random changes.

The paper identifies gamma-power ratio, theta–beta balance, signal unpredictability and complexity as important predictive features. Ablation experiments reported accuracy reductions of 3.21 percentage points without data augmentation, 5.77 points when the

ensemble was replaced by a single algorithm, and 4.96 points when the feature set was reduced from 47 measurements to 20. These findings support the components’ contributions within the reported experiments, although independent testing is needed to establish generalizability.

Responsible AI and Scientific Credibility

The paper describes a responsible AI framework containing 46 modules and more than 1,300 analysis types. It examines fairness, privacy, safety, transparency and robustness, using SHAP and LIME to help explain predictions.

The reported governance score is 0.91 out of 1, with assessments mapped to expectations associated with the EU AI Act, US FDA guidance for medical software and HIPAA privacy rules. This is a research-defined governance assessment, not regulatory approval, formal certification or proof of legal compliance. Patient privacy should likewise not be described as guaranteed without evidence of the specific tests performed. The paper acknowledges that further multi-centre studies are needed.

How Well Does It Work?

In the main evaluation, reported accuracy was 100% for Parkinson’s  disease, 99.02% for epilepsy, 97.67% for autism spectrum disorder, 97.17% for schizophrenia, 94.20% for Alzheimer’s disease, 94.17% for chronic stress and 91.07% for major depression. The corresponding area-under-the-curve (AUC) values ranged from 0.956 to 1.000.

However, a stricter leave-one-out evaluation reported Parkinson’s disease accuracy of 92.4% and schizophrenia accuracy of 91.2%. This difference  demonstrates why evaluation design matters. The 100% Parkinson’s result must be presented alongside the lower result rather than as a universal guarantee of diagnostic accuracy.

The original methods should be checked to confirm participant-level separation, the handling of overlapping and augmented data, and the differences between evaluation procedures. High accuracy alone cannot establish that data leakage was avoided.

Limitations and Future Directions

The study uses data from a single centre, limiting conclusions about performance across different  hospitals and patient populations. EEG signals can vary with recording conditions, medications and other health conditions. The current system addresses classification but does not establish the ability to assess disease severity or distinguish closely related subtypes.

Further research should report detailed evaluation methods, participant numbers, reference standards and uncertainty around performance estimates. Independent replication and prospective, multi-centre validation are essential before clinical reliability can be established.

Future directions include federated learning, portable EEG devices, integration with other clinical information and potential predictive applications, such as identifying signs preceding epileptic seizures. These remain proposed developments rather than established capabilities of the current system.

The Bottom Line

NeuroMCP-Agent combines EEG-based classification with uncertainty estimation, explainability and governance assessments. Its reported results are promising, but the difference between 100% and 92.4% Parkinson’s disease accuracy highlights the importance of rigorous testing and transparent reporting. The governance score is not certification, and privacy claims must be supported by documented evidence.

If independent, multi-centre studies reproduce these findings, NeuroMCP-Agent may eventually contribute to EEG-based screening and clinical decision support. For now,  it remains a promising research system, not a validated replacement for clinical diagnosis.

Background sources: WHO, Global Status Report on Neurology (2025); Global Burden of Disease Study 2021, The Lancet Neurology (2024); WHO, World Mental Health Today (2025). Technical details and performance figures should be verified against the original NeuroMCP-Agent research paper before publication.