The use of hidden semi-markov models in clinical diagnosis maze tasks

In this paper, we investigate the use of hidden semi-Markov models (HSMMs) in analyzing data of human activities, a task commonly referred to as activity recognition. In particular, we use the models to recognize normal and abnormal two-dimensional joystick-generated movements of a cursor, controlled by human users in a computerized clinical maze task. This task - as many other activity recognition tasks - places a lot of emphasis on the duration of states. To model the impact of these durations, we present an extension of HSMMs, called Non-Stationary Hidden Semi Markov Models (NSHSMMs). We compare the performance of HMMs, HSMMs and NSHSMMs in recognizing normal and abnormal activities in the data, revealing the advantages of each method under different conditions. We report the results of applying these methods in analyzing real-world data, from 75 subjects executing clinical diagnosis maze-navigation tasks. For relatively simple activity recognition tasks, both HSMMs and NSHSMMs easily and significantly outperform HMMs. Moreover, the results show that HSMM and NSHSMM successfully differentiate between human subject behaviors. However, in some tasks the NSHSMMs outperform the HSMMs and allow significantly more accurate recognition. These results suggest that semi-Markov models, which explicitly account for durations of activities, may be useful in clinical settings for the evaluation and assessment of patients suffering from various cognitive and mental deficits.