Traditional prediction models face challenges in analyzing time-series electronic health records (EHR). Our research focuses on utilizing Recurrent Neural Networks (RNN), Transformers, and self-supervised learning to achieve precise disease risk prediction using time-series data. Additionally, our approach allows us to identify influential time points and predictive factors, thus advancing the goal of precision medicine.

We also investigate how to enhance novel interpretability mechanisms like Testing with Concept Activation Vectors (TCAV), enabling their broader application in explaining biomedical deep learning models.