Drawing-Based Automatic Dementia Screening Using Gaussian Process Markov Chains

Abstract

Screening tests play an important role for early detection of dementia. Among those widely used screening tests, drawing tests have gained much attention in clinical psychology. Traditional evaluation of drawing tests totally relies on the appearance of drawn picture, but does not consider any time-dependent behaviour. We demonstrated that the processing speed and direction can reflect the decline of cognitive function, and thus may be useful for disease screening. We proposed a model of Gaussian process Markov chains (GPMC) to study the complex associations within the drawing data. Specifically, we modeled the process of drawing in a state-space form, where a drawing state is composed of drawing direction and velocity with consideration of the processing time. For temporal modeling, our scope focused more on discrete-time Markov chains on continuous state space. Because of the short processing time of picture drawing, we applied higher-order of Markov chains to model long-term temporal correlation across drawing states. Gaussian process regression was used for universal function approximation to flexibly infer the state transition function. With Gaussian process prior to the distribution of function space, we could encode high-level function properties such as noisiness, smoothness and periodicity. We also derived an efficient training mechanism for complex Gaussian process regression on bivariate Markov chains. With GPMC, we present an optimal decision rule based on Bayesian decision theory. We applied our proposed method to a drawing test for dementia screening, i.e. interlocking pentagon-drawing test. We tested our models with 256 subjects who are aged from 65 to 95. Finally, comparing to the traditional methods, our models showed remarkable improvement in drawing test for dementia screening.