Predicting Length of Stay in Acute Care Using Day-to-Day Patient Information

Predicting the Length of Stay (LoS) in healthcare settings is a critical task that supports optimized resource allocation and tailored clinical decision-making. Unlike most studies focused on ICU patients, this work targets acute care settings, addressing the unique challenges of irregular and intermittent temporal data. The objective is to develop a model that integrates multimodal day-to-day measurements, combining temporal clinical data with demographic information. We frame LoS prediction as a regression problem, leveraging historical trends to account for irregularities and gaps in the data. A novel approach employing sentence transformers is introduced to generate data embeddings, enabling a detailed representation of patient conditions. This is compared against an LSTM-based model, which has shown promise in similar studies. Preliminary results demonstrate the sentence transformer model's superior performance in capturing data intricacies, paving the way for more accurate and clinically significant predictions.