Synopsis

The CSV can be found here

Inspiration

Mental health disorders are complex and often misdiagnosed due to the subjective nature of psychiatric evaluations. EEG data provides real-time insights into brain activity, but analyzing it manually is challenging due to its size and complexity. We were inspired by the potential of machine learning to improve diagnostic accuracy, making psychiatric assessments more objective and reliable.

What it does

NeuroMind AI is a machine learning model designed to predict major psychiatric disorders using EEG data. By leveraging advanced algorithms, the model can classify mental health conditions such as mood disorders, anxiety, and trauma-related disorders. The goal is to enhance clinicians’ ability to diagnose mental health conditions accurately, reducing bias and improving patient outcomes.

How we built it

Data Collection & Preprocessing

We worked with EEG datasets containing labeled mental health conditions. Missing data was handled through imputation, and categorical variables (such as demographic features) were encoded.

Model Selection & Training

We explored multiple models, including Logistic Regression, Random Forest, CNNs, and Transformers. XGBoost was chosen due to its interpretability and training efficiency. Hyperparameter tuning was performed to optimize classification accuracy.

Model Evaluation

Baseline accuracy for different models: Logistic Regression: 10.6% ± 2.7% Random Forest: 30.9% ± 2.1% CNN: 32.4% Transformer: 30.1%

After fine-tuning, we achieved 39.9% accuracy, outperforming baseline models.

Challenges we ran into

Data Complexity: EEG signals contain noise, making preprocessing crucial. Low Accuracy: Despite tuning, accuracy remained below 50%, highlighting the difficulty of classifying psychiatric disorders. Dataset Limitations: We had to inject additional data from published research to enrich training sets.

Accomplishments that we're proud of

Successfully developed an EEG-based classifier for psychiatric disorders. Improved diagnostic accuracy by 9% over CNNs and Transformers. Integrated external datasets to enhance model generalization. Highlighted the potential of AI in reducing bias in mental health diagnosis.

What we learned

EEG-based machine learning classification is highly challenging due to signal variability. Feature engineering (e.g., extracting PSD features) significantly impacts performance. While deep learning models showed potential, tree-based models like XGBoost provided better interpretability and efficiency. Multiclass classification in mental health requires larger, high-quality datasets.

What's next for NeuroMind AI

Expand Dataset: Incorporate larger and more diverse EEG datasets for better generalization. Optimize Deep Learning Approaches: Explore advanced architectures like attention-based EEG models. Clinical Validation: Partner with hospitals and research institutions to test the model in real-world psychiatric assessments.