Heart data analysis, modelling and application in risk assessment

The heart is a fundamental aspect of the human body. Significant work has been undertaken to better understand the characteristics and mechanisms of this organ in past research. Greater understanding of the heart not only provides advances in medicine but also enables practitioners to better assess the health risk of patients. This thesis approaches the study of the heart from a health informatics perspective. The questions posed in this thesis is whether research is capable of describing and modelling heart data from a statistical perspective, along with exploring techniques to improve the accuracy of clinical risk assessment algorithms that rely on this data. The contributions of this thesis may be grouped into two main areas: statistical analysis, modelling and simulation of heart data; and improved risk assessment accuracy of the Early Warning Score (EWS) algorithm using a quartile-based technique. Statistical analysis of heart data, namely RR intervals, contributes to more informed understanding of the underlying characteristics of the heart and is achieved using null-hypothesis testing through the Anderson-Darling (AD) test statistic. The modelling process of heart data demonstrates methodologies for simulation of this data type, namely individual distribution modelling and normal mixture modelling, and contributes to assessing techniques that are most capable of modelling this type of data. For improved accuracy on the EWS algorithms, a quartiles technique, inspired by anomaly-based intrusion detection systems, is presented which enables customisation of risk score thresholds for each patient defined during a training phase. Simulated heart data is used to evaluate the standard EWS algorithm against the quartile-based approach. The defined metric of accuracy ratio provides quantitative evidence on the accuracy of the standard EWS algorithm in comparison with the proposed quartile based technique. Statistical analysis in this thesis demonstrates that samples of heart data can be described using normal, Weibull, logistic and gamma distribution within the scope of two minute data samples. When there is strong evidence to suggest that RR intervals analysed fits a particular distribution, individual modelling technique is the ideal candidate whilst normal mixture modelling is better suited for long-term modelling, i.e. greater than two minutes of heart data. In comparative evaluation of the standard EWS algorithm and the quartile-based technique using modelled heart data, greater accuracy is demonstrated in the quartiles-based technique for patients whose heart rate is healthy, but outside the normal ranges of the general population.