Preparing for Disease X: Predicting ICU Admissions Using Time Series Forecasting with Decoder-Only Transformer Neural Networks

Abstract

The COVID-19 pandemic has underscored the critical importance of predictive modelling in managing healthcare resources and shaping public health policies. This paper explores the application of advanced Artificial Intelligence (AI) techniques, specifically decoder-only transformer neural networks (DOTNN), in forecasting weekly Intensive Care Unit (ICU) admissions. Our research is driven by the necessity to enhance preparedness for potential future pandemics, referred to as "Disease X", by leveraging large datasets of publicly available information. A prediction model has been developed that incorporates several key indicators, such as new cases, ICU admissions, and testing rates. Our DOTNN architecture, inspired by the Generative Pre-trained Transformer (GPT), focuses on time series forecasting without the necessity for encoder components, thereby streamlining the prediction process. Despite limited data availability, the proposed method can achieve notable accuracy, with Mean Absolute Percentage Error (MAPE) values below 15% for a significant number of predictions. This performance highlights the potential of DOTNNs in forecasting ICU admissions, which is crucial for healthcare planning and resource allocation during pandemics.