High-accuracy detection of early Parkinson's Disease using multiple characteristics of finger movement while typing

Abstract

Parkinson's Disease (PD) is a progressive neurodegenerative movement disease affecting over 6 million people worldwide. Loss of dopamine-producing neurons results in a range of both motor and non-motor symptoms, however there is currently no definitive test for PD by non-specialist clinicians, especially in the early disease stages where the symptoms may be subtle and poorly characterised. This results in a high misdiagnosis rate (up to 25% by non-specialists) and people can have the disease for many years before diagnosis. There is a need for a more accurate, objective means of early detection, ideally one which can be used by individuals in their home setting. In this investigation, keystroke timing information from 103 subjects (comprising 32 with mild PD severity and the remainder non-PD controls) was captured as they typed on a computer keyboard over an extended period and showed that PD affects various characteristics of hand and finger movement and that these can be detected. A novel methodology was used to classify the subjects' disease status, by utilising a combination of many keystroke features which were analysed by an ensemble of machine learning classification models. When applied to two separate participant groups, this approach was able to successfully discriminate between early-PD subjects and controls with 96% sensitivity, 97% specificity and an AUC of 0.98. The technique does not require any specialised equipment or medical supervision, and does not rely on the experience and skill of the practitioner. Regarding more general application, it currently does not incorporate a second cardinal disease symptom, so may not differentiate PD from similar movement-related disorders.

Fig 1. Example of keystroke timings while typing the word ‘GOAD’.

Fig 2. Data collection and processing flow.

Fig 3. The areas of the keyboard included in keystroke capture.

Fig 4. Distribution of participant ages (all participants, both PD and controls).

Fig 5. Machine learning flow.

Fig 6. Area under the curve (AUC) for Group A results.

(The clinician results shown are from Schrag et al. [45]).

Fig 7. The range of Group A classification probability values.

PD and non-PD, showing the 99% confidence interval for each.

Fig 8. Area under the curve (AUC) for Group B results.

Fig 9. The range of Group B classification probability values.

PD and non-PD, showing the 99% confidence interval for each.