Machine Learning Solutions for Patient Monitoring during Anesthesia

Maciag, Tomasz (2019) Machine Learning Solutions for Patient Monitoring during Anesthesia. Master's Thesis / Essay, Human-Machine Communication.

Preview	Text mHMC_2019_MaciagT.pdf Download (2MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (119kB)

Abstract

In an operating room, one of the roles of an anesthesiologist is to track the patient’s vital parameters displayed on two monitors. These have certain thresholds for each measured value which when surpassed trigger an alarm. Such a simplistic approach often leads to situations where most of the warnings are false, hence the credibility of the whole alarming system is low. Moreover, several potentially dangerous complications consist of compound interactions between different variables in prolonged periods of time; therefore, a solution capable of capturing a higher underlying complexity is needed. This study aims to investigate how Machine Learning techniques can be applied to improve the quality of alarms during anesthesia. Two distinct approaches were tested – Complication Detection and Anomaly Detection. In the former, several algorithms were trained to generate improved warnings for hypotension (low blood pressure), while in the latter, more general approach, different methods of detecting anomalous health states were evaluated. We showed that Complication Detection, based on supervised Neural Networks, was the best performing method; however, we did not rule out Anomaly Detection, as it was in overall a more flexible approach that gives more possibilities for future improvements. Finally, we stated that Machine Learning offers promising solutions for patient monitoring, nevertheless due to a phenomenon called the precision-recall tradeoff, it might be difficult to reduce the number of false alarms without a small decrease in the number of detected complications, regardless of the applied technique. The outcome of this study leads to a number of guidelines for future research in this area, including appropriate preprocessing of medical time-series data, comparisons between simple algorithms and more intricate deep learning architectures.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Cnossen, F.
Degree programme:	Human-Machine Communication
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	24 Mar 2019
Last Modified:	25 Mar 2019 12:05
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/19293

Actions (login required)

View Item