Capturing accelerometer outputs in healthy volunteers under normal and simulated-pathological conditions using ML...

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning in Healthcare

Background:

• Wearable devices enable objective physical activity measurement.
• Current activity recognition algorithms are primarily trained on healthy populations.
• Algorithm suitability for clinical scenarios, like altered gait, remains uncertain.

Purpose of the Study:

• To investigate the accuracy of wearable device algorithms trained on healthy data when applied to simulated pathological gait.
• To evaluate if retraining algorithms with simulated pathological data improves classification accuracy.
• To assess the impact of gait alterations on physical activity classification.

Main Methods:

• Healthy participants (n=30) performed nine activities under healthy and simulated-pathological gait conditions.
• Data were collected using wrist-worn MOX accelerometers.
• Multiple machine learning models (SVM, Neural Network, Random Forest, k-NN, Naive Bayes) were trained and tested using both healthy and simulated pathological gait datasets.

Main Results:

• Support Vector Machine (SVM) achieved 98.4% accuracy for activity classification using healthy data.
• Accuracy dropped significantly to 52.8% when the healthy-trained model was tested on simulated pathological gait data.
• Retraining the SVM model with simulated pathological data improved accuracy to 96.7% for classifying abnormal activity.

Conclusions:

• Physical activity recognition algorithms trained on healthy data exhibit decreased accuracy in individuals with altered gait.
• Developing and utilizing classifier algorithms trained on data from specific sub-populations, such as those with simulated pathological gait, is crucial for restoring high accuracy.
• Tailoring algorithms to the target population is essential for reliable wearable sensor data interpretation in clinical settings.