Activity recognition in patients with tremor: Integrating time–length windows for enhanced detection

Pathological tremor significantly impairs daily activities and quality of life, particularly in conditions such as essential tremor and Parkinson's disease. The assessment and evaluation of tremor, along with its evolution with medication dosages, pose a challenging problem. Current clinical evaluation methods are constrained by patient performance, variability, and evaluator's subjectivity. Introducing continuous monitoring in everyday life settings could offer insights by linking tremor occurrences with specific activities. However, existing studies on classifying activities of daily living either exclude tremor patients or rely on fixed-size windows for performance modelling. These approaches often adopt window sizes based on past successful cases without clear consensus on optimal sizing. This study proposes a novel method using a single smartwatch to classify ADLs. By processing time series data with various windowing techniques and applying machine learning models (Support Vector Machines, Random Forest, and Extreme Gradient Boosting), we aim to improve tremor monitoring. This approach offers a more comprehensive and objective evaluation, potentially enhancing treatment strategies. Evaluation metrics including accuracy, precision, recall and F1-score statistics are used to assess the identification abilities of the proposed models. Our results demonstrate significant improvements in activity recognition, highlighting the effectiveness of integrated time-length windows for feature extraction and ADL classification.