Local and personalised modelling for renal medical Decision Support System
Ma, Tian Min
MetadataShow full metadata
The complexity and the dynamics of real-world problems, such as large Health Informatics data processing, require sophisticated methods and tools for building adaptive and knowledge-based intelligent systems. This research developed intelligent systems for Health Informatics, and focuses on those local and personalised modelling which perform better local generalisation over new data. The local models are based on the principles of local learning, where the data is clustered and for each cluster a separate local model is developed and represented as a fuzzy rule as a knowledge representation, either of Takagi-Sugeno, or Zadeh-Mamdani types. The personalised modelling techniques are based on transductive reasoning. They develop individual model for each data vector that takes into account the new input vector location in the space. They are adaptive models, in the sense that input-output pairs of data can be added to the data set continuously. This type of personalised modelling is promising for medical decision support systems where a model for each patient is developed to predict an outcome for this patient and to rank the importance of the clinical variables for them. This thesis presents novel local and personalised modelling and illustrates them on real world medical case studies of renal function evaluation – an important problem of medical decision support. The local and personalised models are compared with statistical, neural network and neural fuzzy global models and show a significant advantage in accuracy and explanation. Two representative problems in clinical medicine have been explored using the framework of local and personalised modelling. In each case, prediction has been made utilising either clinical, laboratory, or a combination of different types of data where appropriate. Systems has been developed for the following circumstances: (1) prediction appertaining to renal function, using data from 178 Australasian patients with advanced chronic kidney disease (computing procedure GFR-DENFIS, GFR-KBNN, GFR-TWNFI); (2) prediction appertaining to patient longevity after the inception of dialysis for end-stage renal failure,using data from 6010 patients randomly sampled from United States facility haemodialysis population (computing procedure DOPPS-TWNFC, DOPPS- TTLSC). The main contribution of this research is to provide immediate and workable methods and tools to augment health care, which are of sufficient accuracy to support good clinical decision-making. Furthermore, this research resulted in technical solutions to the various data modelling problems that exist in health care research. More importantly, personalised modelling developed for renal disease in this research is an adaptive and evolving technique, in which new data sample can be continuously added to the training dataset and subsequently contribute to the learning process of personalised models. The technique of personalised modelling offers a new tool to give a profile for each new individual data sample. Such characteristic makes personalised modelling based methods promising for medical decision system, especially for complex human disease diagnosis and prognosis.