Precision-Recall Adversarial Network (PRAN): A Novel Deep Learning Framework for Improvement of Clinical Binary Classification
Project Summary: Clinical machine learning models demonstrate promise in improving prediction of medical outcomes, but imbalanced class distribution in the training data presents challenges for optimal model performance. Precision and recall are strong metrics that describe the performance of a model on the minority class, but improving both metrics is difficult due to their tradeoff. I propose the Precision-Recall Adversarial Network (PRAN), a deep learning framework in which one model improves recall while another improves precision. I hypothesized that both models would compete and improve both metrics together. This model was tested on four clinical imbalanced datasets from the UCI repository, including classification of hepatitis survival, thyroid disease, breast cancer prognosis, and Parkinson’s Disease. In the first three respective datasets, PRAN and PRAN paired with SMOTE outperformed most selected classifiers in terms of F1-Score. The framework demonstrates potential as a viable option to improve classification performance on imbalanced clinical data.