A system of decision support for medical diagnosis

This thesis presents DAMOCLES, a quantitative modelling approach to medical diagnosis that addresses several shortcomings of existing approaches to computer-assisted diagnosis. In developing this approach, several important epistemological issues are explored, and the nature of necessary anatomical, physiological, pathological and clinical knowledge in the medical domain is analysed in detail. A domain model architecture to represent this knowledge is proposed, the main contribution of which is an inductive method for determining from raw data the form of functions describing the relationships between sets of variables, with no a priori assumptions made about the form of the relationships or the distributions of the data. The representation of these functions contains a contour of nonparametric conditional probability estimates across the range of values of the dependent variable. Experimental work demonstrates the potential for a high degree of accuracy in both the estimate of the form of a function and in the estimates of conditional probabilities. Applying the functions as diagnostic tests, it is shown that sensitivity and specificity in excess of 90% can be achieved, and that if multiple observations over time are used as a sample then sensitivity and specificity can approach 100%. A diagnostic strategy is developed to determine overall domain solutions from a set of functions and observations, and this is shown to outperform the author in a particular diagnostic task within a domain of high dimensionality and interconnectivity. Lastly, it is discussed how various medical heuristics in common use can be added into the diagnostic process, and how a comprehensive medical domain model might be assembled, which would need to include some form of qualitative modelling not addressed in this thesis