PhD defence by Søren Straarup Rasmussen

On Thursday, 26 October, Søren Straarup Rasmussen will defend his PhD thesis " Automatic Interpretation of Vital Signs for Detecting Physiological Deterioration".

Time: 13.00

Place: Building 308, auditorium 11 & zoom:  https://dtudk.zoom.us/meeting/register/u5Isd-irrTwqHdeNVM0Oqve4oXDIlmilpjc_

Please be aware that the PhD defense may be recorded - This will also be informed at the beginning of the PhD defense.

Supervisor: Professor Jakob Eyvind Bardram

Co-supervisor: Professor Christian Sylvest Meyhoff

Professor Eske Kvanner Aasvang

Assessment committee:
Associate Professor Jes Frellsen, DTU Compute
Associate Professor Iain Rice, Birmingham City University
Associate Professor Jørgen Kanters, University of Copenhagen

Chairperson:
Associate Professor Sadasivan Puthusserypady, DTU Health Tech

Abstract:
While technological advances in recent years have increased the use of wireless biosensors for monitoring biological signals for a range of different uses, the majority of hospitalized patients still rely on manual measurements for medical staff to track their physiological condition. However, systems for vital signs monitoring are now entering the market with the aim to improve patient safety and care at hospitals. These systems will change the paradigm of patient monitoring by introducing temporal and timely data in a resolution not previously adapted at general wards. With the increasing volume of biological signals, the one question to answer is: how are these to be analyzed and used to identify early signs of deterioration? By using unambiguous biosensors on patients during their hospitalization, we aim to explore methods for analyzing the signals obtained given the challenges they entail, i.e. obtaining proper labels, working with missing data, and identifying early markers of physiological deterioration.
As much of the data recorded during vital signs monitoring is not easily labeled for supervised learning, we explore the use of a Deep Generative Model to utilize unlabeled and labeled data in a semi-supervised manner. We design and implement a probabilistic framework based on the Variational Auto-Encoder to identify Atrial Fibrillation from over 110,000 Electrocardiography segments with different proportions of labeled and unlabeled data. The results demonstrate the potential of using semi-supervised learning for classification tasks, as it was superior to fully supervised learning with only 5% of the data labeled.
The application of machine and deep learning in health care-related topics is criticized due to the lack of explainability of the models, as parameters that are recognizable in the clinic have a higher potential for compliance by the staff. On the general ward, the nurses follow the directions for care stipulated by different scores in the Early Warning Score (EWS) protocol. Whereas several continuous monitoring systems have implemented the EWS protocol to achieve recognizability from the staff, this only allows for detecting deviations after they have occurred. We addressed this challenge by developing and testing a Multivariate Auto-Regressive model for forecasting in the vital signs time series. The model was evaluated on the ability to forecast 15 minutes into the future, and it showed reasonable predictive accuracy on unseen patients.
The main objective of using continuous vital signs monitoring in the general wards of hospitals is to be able to identify the deteriorating patient earlier than is currently possible. As the definition of deterioration is ambiguous and dependent on many internal and external factors, we explored the use of novelty detection to detect deterioration as large deviations from normality. We used a Kernel Density Estimation (KDE) model to establish a distribution from segments of patient recordings defined to be normal. By assessing new segments in relation to this, a stability index was computed from the likelihood, quantifying the physiological condition of the patient. We showed that severe deviations in the vital signs could be identified several hours prior to identification by staff.
The work presented in this thesis identifies challenges within analyzing continuous vital signs and explores different ways of negotiating these. The work proposes ways for handling the uncertainty present in the labeling of the recordings and for forecasting vital signs using generative models. Furthermore, different methods for identifying the deteriorating patients are implemented and tested against different available endpoints.