Multimodal Deep Learning Framework for Early Sepsis Prediction Using ICU Physiological and Clinical Data

Abstract

Sepsis is a major cause of mortality worldwide, particularly in Intensive Care Units (ICUs), where timely diagnosis and early prediction are essential for improving patient outcomes. Sepsis is a major cause of death in the world and more so in intensive care units (ICUs) where a timely given diagnosis and prompt prognostication are imminent factors in death. Traditional clinical scoring procedures and many machine learning procedures often do not include physiological biomarkers with dynamic temporal changes that determine the development of diseases, which results in limiting their predictive value. To counter this, the current study presents a multitask deep learning network aimed at predicting the development of sepsis based on longitudinal physiological and clinical data that are collected among ICU patients. The suggested architecture 1 combines heterogeneous electronic health record components: vital sign traces, laboratory analysis results and demographic factors into consistently divided time series to forward them to high-level sequencesensitive deep learning models namely Temporal Convolutional Network (TCN) and a coupled Convolutional Neural Network Long Short-Term Memory (CNN-LSTM) architecture. Temporal masking and median-imputed methods are used in order to attenuate the impacts of the skipped entries and varying sequence lengths. Evaluation of the predictive accuracy visually empirically demonstrates that the TCN has reached the highest predictive accuracy of 0.9785, followed by CNN-LSTM with 0.9730 and a gradient boosting ensemble that has reached 0.9690. These results emphasize the substantive benefit that exists as a result of temporal modeling in the improvement of the predictive systems about early sepsis detection.