Abstract
Population aging has intensified the prevalence of falls among older adults, making automatic Fall Detection Systems (FDS) a key component of telemonitoring and remote care. Among wearable-based approaches, inertial sensors, particularly accelerometers, offer an effective and low-cost alternative for continuous monitoring. However, the impact of the selection of the sampling frequency on model performance remains insufficiently explored. This work seeks to determine the sampling rate that best balances accuracy, stability, and computational efficiency in wearable FDS. Five representative algorithms (CNN-LSTM, CNN, LSTM-BN, k-NN, and SVM) were trained and evaluated using the SisFall dataset at 10, 20, 50, and 100 Hz, followed by a multi-stage validation including the real-fall repositories FARSEEING and Free From Falls, as well as a seven-day continuous monitoring test under real-life conditions. The results show that deep learning architectures consistently outperform traditional classifiers, with the CNN-LSTM model at 20 Hz achieving the best balance of accuracy (98.9%), sensitivity (96.7%), and specificity (99.6%), while maintaining stable performance across all validations. The observed consistency indicates that intermediate frequencies, around 20 Hz and down to 10 Hz, provide sufficient temporal resolution to capture fall dynamics while reducing data volume, which translates into more efficient energy usage compared to higher sampling rates. Overall, these findings establish a solid empirical foundation for designing next-generation wearable fall-detection systems that are more autonomous, robust, and sustainable in long-term IoT-based monitoring environments.