Abstract
OBJECTIVE: We aim to propose a novel method of evaluating the degree of rhythmic irregularity during repetitive tasks in Parkinson's disease (PD) by using autocorrelation to extract serial perturbation in the periodicity of body part movements as recorded by objective devices. METHODS: We used publicly distributed sequential joint movement data recorded during a leg agility task or pronation-supination task. The sequences of body part trajectory were processed to extract their short-time autocorrelation (STACF) matrices; the sequences of single task conducted by participants were then divided into two clusters according to their similarity in terms of their STACF representation. The Unified Parkinson's Disease Rating Scale sub-score rated for each task was compared with cluster membership to obtain the area under the curve (AUC) to evaluate the discrimination performance of the clustering. We compared the AUC with those obtained from the clustering of the raw sequence or short-time Fourier transform (STFT). RESULTS: In classifying the pose estimator-based trajectory data of the knee during the leg agility task, the AUC was the highest when the STACF sequence was used for clustering instead of other types of sequences with up to 0.815, being comparable to the results reported in the original analysis of the data using an approach different from ours. In addition, in classifying another dataset of accelerometer-based trajectory data of the wrist during a pronation-supination task, the AUC was again highest up to 0.785 when clustering was performed using the STACF rather than other types of sequence. CONCLUSION: Our autocorrelation-based method achieved a fair performance in detecting sequences with irregular rhythm, suggesting that it might be used as another evaluation strategy that is potentially widely applicable to qualify the disordered rhythm of PD regardless of the kinds of task or the modality of devices, although further refinement is needed.