More than 37 million falls that require medical attention occur every year, mainly affecting the elderly. Besides the natural consequences of falls, most aged adults with a history of falling are likely to develop a fear of falling, leading to a decrease in their mobility level and impacting their overall quality of life. Previous wrist-based datasets revealed limitations such as unrealistic recording set-ups, lack of proper documentation and, most importantly, the absence of elderly people's movements. Therefore, this work proposes a new wrist-based dataset to tackle this problem. With this dataset, exhaustive research is carried out with the low computational FS-1 feature set (maximum, minimum, mean and variance) with various machine learning methods. This work presents an accelerometer-only fall detector streaming data at 50 Hz, using the low computational FS-1 feature set to train a 3NN algorithm with Euclidean distance, with a window size of 9 s. This work had battery and memory limitations in mind. It also developed a learning version that boosts the fall detector's performance over time, achieving no single false positives or false negatives over four days.
Keywords: battery/memory limitations; fall detection; learning version; machine learning methods; wrist-based dataset; wrist-based solution.