Rice Data Challenge 2025

Inspiration Data-Driven Insights for Mental Health: Breaking Barriers, Building Solutions."

"Turning Data into Diagnosis: Smarter Mental Health Solutions.

What it does

Our project aims to improve early detection of psychiatric disorders using machine learning on EEG data. By analyzing PSD and COH data, we can classify disorders more accurately, making mental health diagnostics more accessible and data-driven.

How we built it

Preprocessing: We handled missing values with KNN imputation, normalized features, and applied feature selection using hierarchical clustering.

Modeling: We tested multiple ML models (ExtraTrees, XGBoost, SVM, KNN) and optimized them with Optuna hyperparameter tuning.

Balancing the Dataset: Since some disorders were underrepresented, we used SMOTE to generate synthetic data and balance classes.

Feature Importance Analysis: We used hierarchical clustering based on the Spearman correlation to get the dendogram of features. Based on a cut off threshold we selected the features for the model.

Challenges we ran into

High-Dimensional Data: With over 1,000 EEG features, reducing dimensionality without losing crucial information was difficult.

Class Imbalance: Some disorders had far fewer samples, making classification biased toward majority classes.

Model Interpretability: Understanding how and why the model makes predictions required deep interpretability techniques.

Accomplishments that we're proud of

Successfully built and optimized an EEG-based mental health classifier.

Implemented feature selection techniques that reduced model complexity while retaining accuracy.

What we learned

EEG data preprocessing is crucial for effective modeling.

Feature importance analysis helps in understanding real-world applications.

Automating model selection with TPOT/Optuna saves time and improves performance. Balancing datasets improves classification, but oversampling must be carefully handled to avoid overfitting.

The submission file can be found here