Classifying Parkinson's Disease with Machine Learning: Increasing Objectivity and Robustness

Abstract

Advancements in machine learning (ML) have found widespread success in many healthcare tasks, such as drug discovery, smart health records, and diagnosis. However, ML has yet to introduce an effective diagnostic or monitoring model that uses remotely acquired and non-invasive signals. Such a tool would dramatically democratise healthcare, supplying medical expertise to many that cannot currently access it.

This PhD thesis in computer science aimed to support clinicians that diagnose Parkinson's Disease (PD) by increasing the objectivity and robustness of the classification of people with Parkinson's (PWP). Furthermore, by studying the computational analysis of non-invasive, easily, and remotely acquirable signals, contributions are made towards the remote classification and possibly earlier diagnosis of PD, as well as providing solutions and guidelines that apply to issues faced across many healthcare tasks. This thesis reports on three studies.

I first investigated the classification of PD through the analysis of postural sway, a complex system that concurrently processes visual, vestibular, and somatosensory inputs. Literature that compared the sway of PWP and healthy controls was systematically reviewed, revealing that analyses consisted of significance testing on individual features extracted from postural sway; the use of higher dimensional analyses (multiple features or raw signals) was non-existent. A map of all the features used in literature and their corresponding effect sizes was constructed, indicating that commonly used features were not as effective as others. Sub-analysis also suggested that features were more effective whilst OFF medication (rather than ON medication), that vision did not have a consistent impact, and that a lower sampling rate was associated with a smaller effect size.

Following this, postural sway data gathered at the Canberra Hospital was used to experimentally validate the effect size of these features, as well as the experimental conditions which yielded more effective features. Additionally, the effectiveness of features in an out-of-sample framework was quantified using accuracy and the area under the receiver operating characteristic curve (AUROC) from binary classifiers, which also enabled measuring multiple feature effectiveness. These experiments suggested that projecting samples into higher dimensions resulted in an improved separability between the PWP and healthy controls, with an accuracy of 85% and an AUROC of 96%, at best.

In conjunction to this, I also investigated the classification of PD through sustained vowel phonation, an output of the vocal system. The literature surrounding this has reported exceptional results, with classification accuracies up to 100% on the widely used data sets. However, several evaluation pitfalls were identified that were rarely accounted for, termed `inflationary effects', which inflated the performance across the training, validation, and testing, whilst also acting as barriers to general and skilful learning. An evaluation framework is proposed to eliminate these inflationary effects and more faithfully evaluate generalisation performance. The experiments suggest that the inflationary effects contributed up to 30% of classification accuracy, and that the generalisation performance of these reported models was only slightly better than chance.

The findings of this thesis reveal that, although much progress is left to be made, there are strong grounds for the consideration of using postural sway and sustained phonation as possible bases for the classification of PD in order to aid clinical decision making. Furthermore, this thesis highlights and addresses some of the shortcomings of the use of ML in medical research in general that have limited progress. The thesis offers a guideline for future research into remote diagnosis and monitoring systems that are centred around easily and remotely acquired and non-invasive bio-signals. Show more