Time-Series Forecasting in Healthcare: A Dual-Framework Approach with Latent Class Integration

This study addresses the challenge of forecasting clinical time series data, leveraging the MIMIC-IV dataset to improve predictive accuracy through novel latent class integration techniques. We introduce two enhanced variational recurrent models, CVLSTM and LTVLSTM, that combine disease classification with time-series modeling to capture both local patterns and long-term temporal dependencies. CVLSTM incorporates latent class variables to classify diseases and improve prediction accuracy, while LTVLSTM extends this framework by learning local trends in subsequences and integrating them into the broader temporal structure of the data. Using Root Mean Square Error (RMSE) as the primary evaluation metric, our methods significantly outperform traditional ARIMA, LSTM, and VRNN models, with LTVLSTM achieving the highest accuracy on the MIMIC-IV dataset. These results demonstrate the potential of integrating disease-based priors and hierarchical variational inference for enhancing clinical time series forecasting.