Machine Learning Architectures to Classify Activities of Daily Living and Fall Types from Wearable Accelerometer Data

This study compares different machine learning (ML) approaches to classify motion-based activities of daily living based on accelerometer measurements, focusing on identifying different types of falls. Falls are a significant cause of injuries and deaths in older adults, emphasizing the need for real-time fall detection and alert systems. ML algorithms have shown high accuracy in detecting falls in experimental settings, but their performance in real-world scenarios still needs to be studied. Two publicly available datasets, collected from healthy subjects with smartphone accelerometers, were used for evaluation. Different ML classifiers (Support Vector Machine, Random Forest, CatBoost, and a meta-algorithm) were tested using raw accelerometer data and selected features in the time and frequency domains. Results demonstrated that CatBoost and Random Forest outperformed Support Vector Machine in both datasets when using a carefully chosen set of features instead of the raw data as input. CatBoost, although accurate, showed higher computational costs, making Random Forest a more practical choice for real-world applications. The meta-ML system did not provide a significant advantage over individual algorithms. This research contributes insights into feature selection and computational efficiency in accelerometer data classification. It provides evidence-based recommendations for practitioners working with activities of daily living classification and fall detection, aiming to enhance the use of ML in real-world scenarios and improve the quality of life for individuals requiring monitoring and fall prevention.