Using signature method and deep learning based long short term memory model on time series physiological data to predict sepsis

Abstract

Sepsis is a serious medical condition where the body’s immune system reacts to an infection potentially causing organ failure and death. Globally , sepsis affects over 30 million people annually , resulting in around 6 million deaths. Even as the understanding and treatment improve, sepsis remains a significant cause of mortality and morbidity among critically ill patients. Additionally , the critical need for early detection and intervention continues to be unfulfilled. Numerous risk assessment scores have been created to estimate hospital death rates, aiding in decision-making by forecasting the emergence of serious acute illnesses in high-risk patients. Studies have highlighted how variations in heart rate, respiratory rate, oxygen saturation, and blood pressure can indicate early signs of morbidities in patients. Furthermore, data-driven models employing machine learning have been developed to analyze ICU data and generate these risk assessment scores. With the rise in artificial intelligence and machine-learning techniques, various researchers have proposed algorithms for early sepsis detection. These algorithms identified new predictors of sepsis by analyzing clinical laboratory values and vital signs in NICU and adult ICU patients. This thesis is centered on predicting septicemia likelihood in ICU patients using a machine learning based signature method. Alongside, long short-term memory (LSTM) autoencoders model is developed to forecast neonatal mortality at various intervals throughout a patient's hospital stay. In our research, we discovered that integrating the signature method with LSTM based autoencoders architecture as a self-supervised representation learning, holds promising potential for bedside sepsis detection. The LSTM model learns the representation of the time-series physiological features by training the encoder/decoder and a predictor to achieve the risk-stratified clustering of sepsis patients. We evaluated the performance of our proposed model by comparing with multiple traditional and state-of-the-art deep learning-based techniques for time-series prediction task upon different (i) real world and (ii) public datasets. The real-world dataset, sourced from the iNICU platform, includes data from 22 preterm neonates (10 with sepsis and 12 without sepsis). This dataset encompasses 120 clinical variables and employs blood culture as the definitive standard for diagnosing neonatal sepsis. The larger public dataset contains information on approximately 40,336 ICU patients from two hospitals, with up to 43 clinical variables recorded each hour of the patient's stay in the ICU. This extensive dataset, compiled from the Kaggle repository, adheres to the Sepsis-3 clinical criteria for determining the onset of sepsis. It had time-series data of varying length based on patients’ length of stay. To make advancements and enhance accuracy in predicting sepsis, we must consistently aim to improve. Careful cleaning and selection of data attributes are crucial, laying a solid groundwork for future progress in this field.