The Impact of the Accelerometer Sampling Rate on the Performance of Machine and Deep Learning Models in Wearable...

Area of Science:

• Biomedical Engineering
• Gerontology
• Computer Science

Background:

• Population aging increases fall prevalence, necessitating advanced automatic Fall Detection Systems (FDS).
• Wearable inertial sensors, like accelerometers, are effective for continuous monitoring in telemonitoring and remote care.
• The influence of sampling frequency on wearable FDS performance requires further investigation.

Purpose of the Study:

• To determine the optimal sampling frequency for wearable Fall Detection Systems (FDS).
• To balance accuracy, stability, and computational efficiency in FDS algorithms.
• To provide empirical data for designing sustainable, long-term IoT-based monitoring solutions.

Main Methods:

• Trained and evaluated five algorithms (CNN-LSTM, CNN, LSTM-BN, k-NN, SVM) on the SisFall dataset at 10, 20, 50, and 100 Hz.
• Conducted multi-stage validation using FARSEEING and Free From Falls datasets, plus a seven-day real-life monitoring test.
• Assessed model performance based on accuracy, sensitivity, specificity, and stability.

Main Results:

• Deep learning models (CNN-LSTM, CNN, LSTM-BN) outperformed traditional classifiers (k-NN, SVM).
• The CNN-LSTM model at 20 Hz achieved the highest accuracy (98.9%), sensitivity (96.7%), and specificity (99.6%).
• Intermediate frequencies (10-20 Hz) offer sufficient data for fall detection while reducing computational load and energy consumption.

Conclusions:

• A 20 Hz sampling rate provides an optimal balance for wearable FDS performance and efficiency.
• Intermediate sampling frequencies are suitable for capturing fall dynamics without excessive data.
• Findings support the development of more autonomous, robust, and sustainable wearable FDS for IoT environments.