Gait freezing onset and termination prediction in Parkinson disease using shank sensor acceleration data

Motor impairment Freezing-of-Gait (FoG) is associated with reduced mobility leading to falls usually occurring in advancing Parkinson’s disease (PD). Cueing techniques have been successfully used to alleviate the event. Although, FoG prediction is extensively studied but only paucity of work includes event termination detection. Together with freezing, unfreeze event prediction is advantageous to reduce long-term cueing effects. The focus of carried work is to propose a machine-learning based framework for prediction of freeze and unfreeze events. Based on the Freeze Index (FI), a relabelling scheme is developed to represent four gait states, that are, no-FoG, pre-FoG, FoG, and pre no-FoG using publicly available Daphnet dataset, comprising of 10PD patients. Using nine extracted features from the shank sensor, Random Forest model achieved an accuracy of 95.5% using segmentation window of 2s. While episode-independent GroupKFold (5-fold) and subject-independent evaluation reported accuracy of (91.2 ± 0.9)% and (86.2 ± 4.0)% respectively representing more stricter independence constraints to generalize across episodes and, unseen subjects. Following feature selection, a reduced set of five shank features are obtained using 2s segmentation window and accuracy of 96.2% is achieved. At event level, 197 of 237 annotated FoG episodes are successfully identified at onset, corresponding to per-episode sensitivity of 83.12%. The predicted termination transitions achieved 63.6% cue-deactivation accuracy. These results demonstrate that single shank sensor can provide information for both FoG onset and termination with relatively low sensing and feature-processing requirements. However, framework requires clinical validation and its effectiveness for real-time cueing following prospective validation on larger and more diverse PD population.