DTA

Digital Theses Archive

 

Tesi etd-11052021-234400

Type of thesis
Corsi integrativi di II livello
Author
ESPOSITO, CHRISTIAN
URN
etd-11052021-234400
Title
Non-invasive Blood Pressure Estimation Using Physiological Signals Acquired by the Bedside Monitor
Structure
Cl. Sc. Sociali - Scienze Economiche
Course
SCIENZE ECONOMICHE E MANAGERIALI - SCIENZE ECONOMICHE E MANAGERIALI
Committee
relatore Prof.ssa CHIAROMONTE, FRANCESCA
Presidente Prof. BARONTINI, ROBERTO
Membro Prof. MINA, ANDREA
Membro Prof. FAGIOLO, GIORGIO
Membro Prof. CINQUINI, LINO
Membro Dott. VANDIN, ANDREA
Keywords
  • Blood pressure
  • deep learning
  • signal processing
  • signal quality
Exam session start date
01/12/2021;
Availability
completa
Abstract
Blood pressure (BP) is one of the most common vital signs tracked in critically-ill patients and is used as an indicator of cardiovascular status and is often a target of treatment. Continuous monitoring of blood pressure is achieved through invasive insertion of a catheter directly into an artery. In turn, there are risks associated with these arterial lines that include infection, thrombosis, embolization, discomfort for the patient, and therefore they are not used for prolonged periods of time. More commonly, blood pressure is measured at periodic time intervals using an automated blood pressure cuff, however, this method lacks the temporal resolution to optimally monitor critically-ill patients whose physiological state can deteriorate quickly. Therefore, developing a method to non-invasivley estimate BP using machine learning would alleviate many of the complications associated with invasive recordings. In this study, using a large waveform database acquired from critically- ill children as part of routine care, a cohort of 701 patients, consisting of roughly 1.4 million 60 second intervals, was identified. Next, a deep convolutional neural network with a modified WaveNet architecture was trained to estimate systolic and diastolic blood pressure from one-minute intervals of ECG and PPG waveforms. Two models were trained, a general patient agnostic model and a patient-specific model. Our analysis demonstrates major differences in waveform characteristics within this patient population, and improved performance when the model can learn from patients own vital signals. Model results are discussed in the context of application in a Pediatric Critical Care Unit and the unique challenges this poses.
Files