Research and projects

Summer 2025 REU Blog

We developed a transparent, interpretable linear classifier to predict breast cancer diagnosis using the UCI Breast Cancer Wisconsin Diagnostic dataset. The model is designed to minimize misclassification using linear programming techniques, with a focus on simplicity, interpretability, and clinical relevance.

Objectives

To classify tumors as malignant or benign using a linear decision boundary that minimizes diagnostic error. The project emphasizes interpretability, relying on a small number of key features.

Results

Using five biologically meaningful features (e.g., area and concave points), the classifier achieved 91.7% accuracy on the training set. Sensitivity was 86.3% (malignant detection), and specificity was 95.0%. The model is tunable for clinical use and can be extended to include more features for improved accuracy.

We analyzed resting-state EEG data from 88 participants to identify neural patterns that differentiate Alzheimer's disease, frontotemporal dementia, and healthy individuals. Our goal was to uncover reliable spectral markers of neurodegeneration.

Tools & Technologies: Python, MNE-Python, NumPy, Pandas, Matplotlib, SciPy, EEG data (OpenNeuro ds004504), Support Vector Machines, PCA

Objectives

To examine differences in EEG power and signal variability across diagnostic groups, investigate gender effects on EEG signals, and test whether machine learning can classify individuals based on spectral features.

Results

Participants with Alzheimer's disease showed reduced alpha power in the back of the brain and elevated theta in frontal regions. Individuals with frontotemporal dementia exhibited greater beta and gamma variability in frontal-temporal areas. A support vector machine classifier achieved up to 76.9% accuracy in distinguishing Alzheimer’s disease from healthy controls. Gender showed minimal influence.