Abstract:
Intensive Care Units (ICUs) house vast repositories of patient data, particularly vital signs recorded as time series data, offering invaluable insights into patient health dynamics. However, the sheer volume and complexity of this data render manual analysis impractical, necessitating advanced computational approaches for meaningful extraction and pattern recognition. Traditional statistical methods fall short in uncovering nuanced temporal patterns essential for timely intervention and improved patient outcomes. In response, this project endeavors to harness the power of deep neural network-based language models, specifically a BERT-based transformer trained for Masked Language Modelling (MLM) and Next Sentence Prediction (NSP) tasks. By training and publiclyreleasing unsupervised pre-trained embeddings for vital signs data across varying temporal resolu-tions—5 minutes, 10 minutes, 15 minutes, 30 minutes, and 1 hour—the project aims to constructcomprehensive representations of time series data. Furthermore, the study seeks to facilitate resolution transfer of pre-trained models and engage in temporal pattern analysis using techniques such as sequence mining, recurrent neural networks, and attention mechanisms. These efforts are poised to uncover subtle temporal patterns preceding medical events or diagnoses, thereby enhancing healthcare decision-making and patient care strategies in pediatric ICUs
