A Novel Generic Higher-order TSK Fuzzy Model for Prediction and Applications for Medical Decision Support

This paper introduces a higher-order Takagi-SugenoKang (TSK) neuro-fuzzy inference system and its applications in medical decision support systems. Different from most TSK fuzzy systems that utilize first-order TSK type fuzzy rules, the proposed system is composed of higherorder TKS fuzzy rules that have functions in their consequent parts of the following type: y = b0 x1 b1 x2 ...xp . The type of the non-linear function has been chosen based on the rationale that it is well established in the medical practice, for example the MDRD formula and many other formulas for Glomerular Filtration Rate (GFR) evaluation used in the area of renal research and practice [8]. The proposed approach consists of three steps: (1) Apply Fuzzy C-means clustering to partition the input space; (2) Initialize a higher-order TSK fuzzy rule set according to the clustering results. For each fuzzy rule, a non-linear function of the defined type (e.g. the MDRD formula) is used in the consequent part; (3) Train the system on training data with the use of the steepest descent algorithm (Back-propagation learning algorithm) to optimize the parameters of the fuzzy rules. We illustrate the performance of the proposed system on a real case study data for GFR evaluation. For a comparison, we also present the predicted results on the same data by three other models: (i) the MDRD formula; (ii) the MLP network; and (iii) the Adaptive Neuro-Fuzzy Inference System (ANFIS) [2, 4]. The proposed system outperforms the rest as it takes into account some existing knowledge the type of the function already developed in the past and the variables used in it, along with applying a new method for training the system.