Applying Deep Learning, Machine Learning, and Image Transformation to Forecast Septic Shock with the Integration of Multiple Types of Biomarkers Including Heart Rate Variability

The definition of sepsis was re-defined in 2016 to replace the previous “severe sepsis” to increase the prognostic accuracy. However, the incidence and mortality of sepsis did not change significantly since the last decade in Taiwan and the world. The challenge can partly explain the difficulty of predicting the outcomes of therapy for patients with clinically suspected sepsis, despite the current advancement of machine learning and deep learning. Furthermore, researchers are unsatisfactory of merely predicting inevitable outcomes such as mortality and turning to prognosticate patient-centered outcomes such as shock development and responsiveness to therapy. With non-invasive monitoring widely available and affordable during the past few years, hemodynamics and the heart rate variability monitoring becomes possible in front-line health-care settings such as emergency departments. Sepsis, as a disorder widely influence the immunologic and cardiovascular system vias the sympathetic nervous system in the human body, could be monitored via different analysis domains of the heart rate variability motoring. Many researchers have proved the concepts of early detection of septic shock by monitoring the heart rate variability; however, larger studies are still merited to translate the concept to the bedside further. In the meantime, many artificial intelligence and machine learning algorithms, such as recursive partitioning algorithms, support vector machines, and deep learning, are now well developed and can be easily implemented. However, to take advantage of the deep learning algorithms, many researchers now propose utilizing image transformation before applying these algorithms. Furthermore, different fine-tuninig methods such as imbalance data management, modeling in subgroups, and incorporation with unsupervised learning are believed to improve the performance of these prediction models. In this study, we plan to start a three-part serial study to develop a forecast model of sepsis: 1) a 14-year comprehensive longitudinal retrospective sepsis cohort from the Chang Gung Research Database in all branches in Chang Gung Memorial Hospital to compare the effectiveness of the current clinical practice and biomarkers; 2) a prospective one thousand heart rate variability of sepsis (1KHS) project to evaluate the performance of heart rate variability parameters, focused blood-derived biomarkers, hemodynamics and echocardiography via point-of-care ultrasound to validate the importance of time course in these biomarker in forecasting the patient-centered outcomes including development of sepsis and septic shock and responsiveness to therapy during hospitalization; and 3) To develop different machine learning and deep learning-based forecast models based on different subgroups of patients, including stages of sepsis and phenotypes obtained via unsupervised learning. We anticipate this project to initiate a comprehensive understanding of the interaction between sympathetic nerve system, inflammation, and sepsis progression for different subgroups of patients. We also anticipate this project to bring us more knowledge and experience in practicing machine learning and deep learning-based modeling in critical care and emergency medicine. With the advancement in forecasting the responsive of therapy in treating sepsis, a disorder that large causes mortality and morbidity among patients with all kinds of comorbidities, we believe could bring us a more healthy and productive future.

Project ID:PC10907-1725
External Project ID:MOST109-2314-B182-036