Deep Learning for Imbalanced Time Series Clinical Data

Technologies Used: TensorFlow, Deep Learning, Hyperparameter Tuning, Transformer SMOTE

Project Overview

This project focuses on improving the performance of classifiers on imbalanced time series clinical data using a deep learning approach, integrating advanced techniques to handle data imbalance.

Key Highlights:

• Time Series Classification: Developed 10 classification models, experimenting with various architectures, including Simplified Recurrent Neural Networks (RNNs) with Echo State Cells and Transformers.
• Feature Selection: Incorporated Random Forest for feature selection to improve model efficiency and performance.
• Imbalance Handling: Utilized SMOTE and Borderline SMOTE to address data imbalance, achieving a 49% reduction in data skew.
• Model Performance: Achieved a Test AUC of 95% after optimizing model hyperparameters through comprehensive tuning.

Tools and Techniques:

• TensorFlow: Deep learning framework for building and training the classification models.
• SMOTE: Applied SMOTE and Borderline SMOTE for synthetic oversampling of the minority class.