Enhancing Clinical Time Series Forecasting with Latent Class Integration in Variational Recurrent Neural Networks

EasyChair Preprint 15468

Abstract

Forecasting clinical time series data plays a vital role in healthcare by facilitating early interventions and enhancing patient outcomes. Traditional approaches such as linear and logistic regression, and Recurrent Neural Networks (RNN) have been extensively explored. In addition, generative models like Variational Autoencoders (VAE) have been utilized to handle uncertainty and variability in time series data. However, these methods face challenges in capturing complex temporal dependencies and disease-specific characteristics. We propose two novel variational recurrent neural networks (VRNN) based methods, incorporating patient similarity (VRNN-I+) and latent disease classification (VRNN-II+), to enhance predictive performance in clinical forecasting. The first approach (VRNN-I+) enhances the VRNN model by incorporating temporal data from similar patients as additional domain knowledge, aiming to improve the model’s ability to predict patient outcomes. The second approach(VRNN-II+) transitions from using a standard Recurrent Neural Network (RNN) to a Long Short-Term Memory (LSTM) network, while introducing the disease class as a hidden latent variable within the model to capture complex dependencies in the data. The dataset used in this study is derived from Medical Information Mart for Intensive Care-IV (MIMIC-IV), which provided more comprehensive and up-to-date patient records compared to the widely used MIMIC-III dataset. The preprocessing steps were also described accompanying the dataset. To evaluate the effectiveness of incorporating supplementary information for time series prediction, we used the root mean square error (RMSE) as a metric. We selected three subjects to assess the one-step-ahead prediction accuracy, and our results clearly demonstrate that incorporating additional domain knowledge about the patients significantly improves the accuracy of the VRNN (by 15.4%) for clinical time series forecasting.

Keyphrases: VRNN, VRNN-I+, VRNN-II+

@booklet{EasyChair:15468,
  author    = {Ruiying Chen and Qingang Tang},
  title     = {Enhancing Clinical Time Series Forecasting with Latent Class Integration in Variational Recurrent Neural Networks},
  howpublished = {EasyChair Preprint 15468},
  year      = {EasyChair, 2024}}