Abstract : This thesis presents a new approach of data fusion and applies these notions to the modelisation of probabilistic diagnosis in telemedicine. Our contribution is a new definition of qualified gain in a data fusion process, and an application of dynamic bayesian network to telemedicine diagnosis. The final goal is to remotely regulate the physiological state of a patient.
A first study of the field has shown basic data fusion concepts : data fusion process and qualified gain. Structures and types of data sources and results has emerged during this study too. This approach forms a general framework for the second part of this thesis : modelizing and medical diagnosis for a telemedicine application. This is a typical problem where multiple, uncertain and heterogeneous data sources are needed. The Diatelicproject's goal is to assist kidney disease people at home by monitoring their hydration rate. Physiological data are uncertain, noisy and heterogeneous.
Dynamic bayesian networks are used to modelize causal dependancies, which are typical of medical knowledge. They can deal with uncertainty by using the strong probabilistic formalism. A bayesian network model allows us to do efficient data fusion, in particular, by maximizing the certainty gain, i.e. by detecting hydration rate problems with the highest confidence. For this purpose, we are implemented a bayesian engine, to deal with our experiments. The Diatelic (v3) has been implemented with it.
The physician is able to take the right decision using our data fusion process, because he/she has a precision estimation of the hydration rate of the patient. Health state of the patient could be regulated through the use of this system. New problems have arise during this PhD thesis work: on-line models adaptation, quantifying the data fusion gain and dealing with multiple time-scale bayesian networks.