Statistical Pattern Recognition Techniques for Early Diagnosis of Diabetic Neuropathy by Posturographic Data

Introduction

Neuropathy affects up to 50% of diabetic patients showing different clinical features. Peripheral Neuropathy (PN) is one of most prevalent types of diabetic neuropathy. An early diagnosis in asymptomatic patients is useful in order to make the patients aware of their condition and to activate educational programs oriented to encourage some changes in lifestyle. Moreover, effective treatments are available to prevent further complications. Nerve conduction studies are the gold standard for diagnosing PN but these tests can be challenging for both the patient and the testing physician, hence they are typically prescribed in the presence of strong clues for neuropathy. These observations suggest the need of new, easy, and reliable tools for the early diagnosis of PN.

The loss of sensory perception secondary to PN has a markedly detrimental effect on postural stability during stance and gait. Evaluation of postural steadiness is usually based on the interpretation of centre-of-pressure (COP) measures using a dynamometric platform. High correlations have been found in previous works between the severity of neuropathy and the COP measures. Hence it is sensible to assume that posturographic data based on COP measures can be used as input to statistical pattern recognition techniques to automatically build a diagnosis model. An approach using classical Linear Discriminant Analysis (LDA) has been recently proposed in Fioretti, Scocco, and Ladislao (2010).

In this chapter we survey the main features of the cited paper, by discussing the issues related to the application of such parametric technique for PN diagnosis. Next, we compare this approach with a more general, non-parametric methodology. In particular, the use of the supervised learning algorithm called Bayes Vector Quantizer (BVQ) is discussed. BVQ is applied to nearest neighbor types classifiers, resulting in a simple and cheap classification rule. Its roots in statistical pattern recognition, in particular in Bayes decision theory for the minimization of average misclassification risk, guarantee higher robustness and accuracy performance than traditional nearest neighbor, and the ability to deal with asymmetric classification costs and unbalanced data.

In the following Sections we first introduce both the application domain and some Statistical Pattern Recognition concepts and techniques used in this Chapter. Afterwards, the BVQ algorithm is described as well as the experimental methodology we use for designing a 3-class classification model. Results and future research directions are discussed in the remaining of the Chapter.