Model development to improve patient safety in the operating theatre

Abstract

This project considers the need to use machine learning for supporting anaesthesiologists to predict and detect patient issues. Anaesthesiologists play a vital role within medical care and, especially in South Africa, are involved in nearly all medical care practices. This project is the first of its kind to look holistically at the entire anaesthesiology process, where previous papers have aimed at controlling and investigating only a small portion of the process. It was seen that in South Africa hospitals do not capture live patient data electronically, but rather on paper format. After testing and considering live patient data it was opted to construct an artificial data set as to take the sensitivity of the data into account and aim for a higher model accuracy. An artificial patient data set was constructed using interviews, medical knowledge available to the masses and intuition. This data set was described in detail and the deep complex nature of interrelations of the different variables were highlighted. The data set consisted of a 1000 patients, 500 male patients, 500 female patients, age distributions between 20 and 80 years old, patient heights in metres, patient weights in kilograms, heart rates in beats per minute and lastly, systolic and diastolic blood pressures in millimetres mercury. The data set was analysed by a number of machine learning algorithms and it was found that: J48 decision tree achieved a prediction accuracy of 98.9%, logistic regression 97.8%, k-nearest neighbour 98.3% and lastly, neural networks obtained a 99.7% accuracy. Validation and verification was done via the J48’s decision tree and the models were proven to be fit for use and accurate. From the data it could be seen that future projects that would aim to use machine learning in the pre-,intra- and post-operative care sections of anaesthesiology; that they would have to gather a large data set as to make the models more accurate. Unlike other projects that aim to control the amount of anaesthesia or predict the patient risk beforehand, this project has proven that it is possible to continuously predict patient’s current health status while the operation is under way. The report concluded by stating that the neural networks can be used as a second opinion to classify the patient’s current health status and will be run live with current patient information.