AbstractTime series analysis of physiological data could potentially help healthcare professionals in decision making and intervening dangerous clinical events. Time series analysis of physiological data have been mostly limited to classification based approaches. This might be where only a single point or a class is predicted. Classification based approaches are useful in themselves but they are unable to provide any detail and consequently any insight in to a clinical event. In contrast to this, accurate forecasting of physiological time series data is an open challenge and this could help to provide more details regarding the conditions that might lead to a dangerous clinical event. The few existing works that forecast physiological data have not considered a comparison of forecast strategies and data approaches. Researchers have previously found for other disciplines such comparisons can help in building an efficient model. Moreover, the forecast outcome was not applied to any medical applications. Machine learning algorithms and methodologies have received little attention and are yet to be used in a detailed exploration of time series forecasting for physiological data. This thesis aims to explore the scope of time series forecasting of physiological data for the early intervention of dangerous clinical events (hypotension and bradycardia) through machine learning. Different forecasting strategies and data approaches are considered here and analysed in combination with different machine learning algorithms to find the best. Forecast models are built to explore the scope of different machine learning techniques for the intervention of dangerous clinical events. Patients' physiological time series data (blood pressure and heart rate) are used from the MIMIC II and III databases. The findings of this study appear to demonstrate that forecasting of physiological time series data can be used for the intervention of dangerous clinical events. It
is found that MIMO and DIRMO are the two best strategies and the multivariate based approaches appear to outperform the univariate approaches in forecasting physiological data. The work then goes on to show that supervised machine learning for regression appears to predict events with greater accuracy than without regression. It is also observed that the forecast model with a regression algorithm performs better when forecasting in the gap window as well as the target window. This is only true for a gap window of up to 30 minutes. A forecast model with no gap window forecasts events with more than 99 per cent accuracy. All these scenarios could potentially be informative and useful to health care professionals, forecast model developers and researchers. Furthermore, the availability of the forecast regression data could be a further level of potentially useful information to a healthcare professional.
|Date of Award||Aug 2019|
|Supervisor||John Chiverton (Supervisor) & Branislav Vuksanovic (Supervisor)|