Applied Mathematics and Computation 311 (2017) 22–28 Contents lists available at ScienceDirect Applied Mathematics and Computation journal homepage: www.elsevier.com/locate/amc Performance of small-world feedforward neural networks for the diagnosis of diabetes Okan Erkaymaz a , Mahmut Ozer b , Matjaž Perc c,d,∗ a Department of Computer Engineering, Bulent Ecevit University, Zonguldak, Turkey b Department of Electrical & Electronics Engineering, Bulent Ecevit University, Zonguldak, Turkey c Department of Physics, Faculty of Natural Sciences and Mathematics, University of Maribor, Koroška cesta 160, SI-2000 Maribor, Slovenia d Center for Applied Mathematics and Theoretical Physics, University of Maribor, Mladinska 3, SI-2000 Maribor, Slovenia a r t i c l e i n f o Keywords: Diabetes Small-world network Feedforward neural network Rewiring Newman–Watts model Watts–Strogatz model a b s t r a c t We investigate the performance of two different small-world feedforward neural networks for the diagnosis of diabetes. We use the Pima Indians Diabetic Dataset as input. We have previously shown than the Watts–Strogatz small-world feedforward neural network de- livers a better classification performance than conventional feedforward neural networks. Here, we compare this performance further with the one delivered by the Newman–Watts small-world feedforward neural network, and we show that the latter is better still. More- over, we show that Newman–Watts small-world feedforward neural networks yield the highest output correlation as well as the best output error parameters. © 2017 Elsevier Inc. All rights reserved. 1. Introduction Diabetes is a very common health problem of the modern life, spreading rapidly in the world due to the change of nutritional habits [1]. Although type-1, type-2 and gestational diabetes are all common, especially type 2 diabetes mellitus causes significant morbidity and mortality [2]. Therefore, its early detection is of vital importance. Since some forms of diabetes result in a worldwide epidemic that has made it one of the most serious health problem faced by the humankind [3], enormous efforts have been devoted to its early diagnosis and treatment. Expert systems and artificial intelligence techniques are widely used to aid the diagnosis of diabetes. In this context, the Artificial Neural Network (ANN), originally inspired from real biological networks, has been preferred due to its high classification capability [4]. The architecture of the ANN enables users to construct different types of networks such as feedforward, recurrent and competitive [5]. Among them, a feedforward ANN (FFANN) stands out with its remarkable com- putational speed [4]. In this context, the FFANN has been proved to be an efficient intelligent system for the diagnosis of diabetes [6–11]. Temurtas et al. [12] used a multilayer neural network (MLNN) structure for the diagnosis of Pima Indians diabetes and found that the classification accuracy of MLNN trained by the Levenberg–Marquardt algorithm was better than that of conventional neural networks. Moreover, Wang et al. [2] have developed and evaluated an effective classification ap- proach by means of ANN to identify those at high risk of type-2 diabetes mellitus without biochemical parameters. Soltani and Jafarian [13] used probabilistic ANN (PNN) for diagnosis of diabetes with type-2. ∗ Corresponding author at: Department of Physics, Faculty of Natural Sciences and Mathematics, University of Maribor, Koroška cesta 160, SI-2000 Mari- bor, Slovenia. E-mail addresses: matjaz.perc@uni-mb.si, matjaz.perc@gmail.com (M. Perc). http://dx.doi.org/10.1016/j.amc.2017.05.010 0096-3003/© 2017 Elsevier Inc. All rights reserved.