Effects of Sampling Frequency on Human Activity Recognition with Machine Learning Aiming at Clinical Applications

Area of Science:

• Biomedical Engineering
• Digital Health
• Wearable Technology

Background:

• Human activity recognition using wearable accelerometers is a key digital biomarker for disease assessment.
• Minimizing data volume is crucial for long-term clinical monitoring to reduce power, computation, and communication demands.

Purpose of the Study:

• To identify the minimum sampling frequency for accurate human activity recognition.
• To optimize data collection for long-term wearable-based clinical monitoring.

Main Methods:

• Thirty healthy participants wore nine-axis accelerometers at five body locations.
• Activity recognition models were trained and tested using data sampled at frequencies including 100, 50, 25, 20, 10, and 1 Hz.
• Data from the non-dominant wrist and chest were analyzed.

Main Results:

• Reducing sampling frequency to 10 Hz did not significantly impact recognition accuracy at the wrist or chest.
• A significant decrease in accuracy was observed when the frequency was lowered to 1 Hz, especially for activities like teeth brushing.
• 10 Hz sampling frequency effectively balances recognition accuracy with data volume reduction.

Conclusions:

• A 10 Hz sampling frequency is sufficient for maintaining human activity recognition accuracy in clinical settings.
• Optimized sampling frequencies facilitate long-term patient monitoring and the development of smaller clinical devices.
• This approach supports efficient digital biomarker development for disease severity assessment.