Enhancing Mortality Prediction in ICU: A Deep Learning Framework for Multivariate Time Series Classification

Relying on human diagnosis is potentially prone to errors, time-consuming and leading to treatment mistakes. Early Intensive Care Unit (ICU) mortality prediction is important for determining which patients are most likely to die and for offering appropriate interventions that will preserve their lives. However, conventional diagnostic approaches often face challenges in fully harnessing the extensive information contained within Electronic Health Record (EHR) data, particularly in capturing the intricate temporal dynamics of patient health. In this research paper, a hybrid model leverages the power of Temporal Convolutional Networks (TCN) in conjunction with Bidirectional Gated Recurrent Units (BiGRU) to effectively capture the complex temporal dependencies present in ICU time series data, before constructing the model, the data is pre-processed by cutting-edge Generative Adversarial Network (GAN)-based imputation techniques to handle missing values within the EHR data. Through this comprehensive framework, the goal is to significantly enhance mortality prediction accuracy, empower clinicians with actionable insights into patient prognosis, and ultimately improve patient outcomes in critical care settings.