Unsupervised pre-trained vitals signals embeddings for fine-tuned prognostication: a multi-resolution approach

Abstract

The project introduces an innovative method for predicting shock in ICU patients by leveraging unsupervised pre-trained embeddings from vital signs. Utilizing the extensive eICU and MIMIC datasets, the research focuses on four critical variables: heart rate, systolic blood pressure, respiration, and SpO2. These embeddings are generated using a BERT-based model, adapted for the unique characteristics of time-series vital signs data. The approach employs a multi-resolution technique to capture both high-frequency fluctuations and long-term trends in patient data. The model is then fine-tuned to specifically identify early indicators of shock in ICU patients. The study’s findings demonstrate the model’s effectiveness in accurately predicting shock, with an emphasis on early detection to improve patient outcomes. By integrating unsupervised learning with deep neural networks, this research offers a novel approach to critical care, emphasizing predictive accuracy and timely intervention. The model’s ability to generalize across different patient demographics and clinical settings marks a significant advancement in the field of medical prognostication and AI-driven healthcare solutions.