Gait partitioning methods in Parkinson's disease patients with motor fluctuations: A comparative analysis

Measuring gait quality of patients through wearable sensors has become an useful tool to assist in drug administration. Several methods, based on inertial sensors placed on lower limbs, have been proposed over the last decade to estimate clinical meaningful parameters such as gait phase distribution. Application of those algorithms for monitoring gait abnormalities would have a tremendous potential in patients with Parkinson's disease. However, their applicability to patients with severe gait impairment has not been fully tested and compared. In the present study, we conducted a comparative analysis of gait phase detection methods applied to patients with Parkinson's disease both in OFF and ON levodopa conditions. We compared gait partitioning performance of three already proposed algorithms based on a threshold method and a novel Hidden Markov Model approach. Fourteen subjects with idiopathic PD have been enrolled in this study, and evaluated twice during the same day: both in OFF and in ON conditions. All patients have performed three walking tasks along a 20 m walkway in with four Inertial Measured Units placed on shanks and feet. The sagittal angular velocity of all segments and the linear acceleration of feet have been gathered to evaluate stance and swing phases. Force resistive sensors used as foot-switches have been placed under the feet to estimate the reference gait phase sequence. The goodness (G) of methods was evaluated through the Receiver Operating Characteristic. The results provided an optimum goodness for all examined methods (0 < G < 0.25). The best performance has been achieved with the Hidden Model Markov both in OFF (G=0.01) and ON (G=0.01) levodopa conditions. Our results encourage the applicability of HMM in developments of wearable systems for daily monitoring in patients with motor fluctuations.