Discovering the Typing Behaviour of Parkinson's Patients Using Topic Models

Sensing health-related behaviours in an unobtrusive, ubiquitous and cost-effective manner carries significant benefits to healthcare and patient management. In this paper, we focus on detecting typing behaviour that is characteristic of patients suffering from Parkinson’s disease. We consider typing data obtained from subjects with and without Parkinson’s, and we present a framework based on topic models that determines the differing behaviours between these two groups based on the key hold time. By learning a topic model on each group separately and measuring the dissimilarity between topic distributions, we are able to identify particular topics that emerge in Parkinson’s patients and have low probability for the control group, demonstrating a clear shift in terms of key stroke duration. Our results further support the utilisation of key stroke logs for the early onset detection of Parkinson’s disease, while the method presented is straightforwardly generalisable to similar applications.