Using echo state networks for classification: A case study in Parkinson's disease diagnosis

Highlights

• Early diagnosis of Parkinson's disease is essential to effective treatment.

• Echo state networks facilitate classification from simple movement tasks.

• Rapid fitting of echo state networks proves advantageous for applied systems.

• Appropriate data pre-processing enhances classification accuracy.

Abstract

Despite having notable advantages over established machine learning methods for time series analysis, reservoir computing methods, such as echo state networks (ESNs), have yet to be widely used for practical data mining applications. In this paper, we address this deficit with a case study that demonstrates how ESNs can be trained to predict disease labels when stimulated with movement data. Since there has been relatively little prior research into using ESNs for classification, we also consider a number of different approaches for realising input–output mappings. Our results show that ESNs can carry out effective classification and are competitive with existing approaches that have significantly longer training times, in addition to performing similarly with models employing conventional feature extraction strategies that require expert domain knowledge. This suggests that ESNs may prove beneficial in situations where predictive models must be trained rapidly and without the benefit of domain knowledge, for example on high-dimensional data produced by wearable medical technologies. This application area is emphasized with a case study of Parkinson's disease patients who have been recorded by wearable sensors while performing basic movement tasks.

Introduction

Reservoir computing is a general approach for modelling complex dynamical systems using a large recurrent neural network (RNN), with only the network output weights being trained [1]. Echo state networks (ESNs) are a well known implementation where the output connections are fitted using simple ordinary least-squares regression [2]. Owing to this, ESNs are significantly faster and more scalable than many existing more complex machine learning approaches and are ideally suited for time series analysis. Nevertheless, despite this significant advantage, they have yet to be widely used in data mining applications.

Many medical applications have a need for predictive models that can capture the complexity of biological disease pathways to facilitate personalised healthcare. A good example is the Parkinson's disease (PD) case study considered in this paper. PD is a debilitating progressive neurodegenerative disease that presents with a broad spectrum of movement disorders, which even expert clinicians can find challenging to characterise and discriminate from other related diseases [3]. Wearable sensors can provide significant benefits to patient care by objectively measuring movement disorders in high resolution and therefore help monitor disease progress and their use is becoming increasingly widespread [4].

ESNs, with their ability to model dynamical processes, would seem like a sensible candidate for modelling such data and provide two primary benefits. The first is that they can directly model the raw time series to identify any patterns in the underlying dynamics of the signal that conventional feature extraction techniques may miss. Their second advantage is their rapid training speed, resulting from having a closed form solution. This is an important consideration for applied predictive modelling, owing to the need to train and evaluate candidate models on a range of data sets when performing model selection and evaluation.

In this paper, we consider how ESNs can be applied to the problem of diagnosing PD from movement data of the kind that might be collected using wearable accelerometers. Since there has been little existing work in this area, we focus on exploring the key issue of how inputs and outputs can be mapped to the ESN methodology, and how this affects the predictive accuracy of the model. One aspect that is investigated is whether to segment the data before inputting into the model. This has ramifications for subsequent work on analysis of data recorded from wearable sensors by facilitating simpler processing and analysis at the cost of adding more design time [5]. To evaluate the practicality of the resulting network, a two-fold comparison is performed. First, ESN classifiers are compared to models built on summary features derived using the guidance of an expert in movement disorders, in order to establish whether ESNs offer a more flexible alternative without compromising on accuracy. The second comparison is against previous attempts on this data set, highlighting the ability of ESNs to rapidly fit an accurate model comparable with those produced from complex optimisation routines requiring significantly longer computational time.

The rest of the paper is organised as follows: Section 2 provides details of ESNs and previous applications to both classification tasks and medical problems in general, and Section 3 details the data collection process of the Parkinson's disease movement data. The experimental methodology is laid out in Section 4, while Section 5 presents the results. Finally, Section 6 concludes.