Identification of optimal classifier and sensor placement for fall risk classification using IMU-based gait data

Fall risk can be classified using gait data collected from inertial measurement units (IMU). However, the optimal classifier and placement of IMU sensors to maximize classification performance have not yet been suggested.

Is there an optimal IMU application strategy that yields the highest accuracy for classifying fall risk during gait?

Ninety-three community-dwelling older adults were grouped into low or high risk of a fall. Then, kinematic data were acquired with 10 IMU sensors placed on body segments when they walked 10 m. The mean and variance of linear acceleration and angular velocity of IMU data were used as input features. We compared the performance of models trained using support vector machine (SVM), decision tree (DT), random forest (RF), K-nearest neighbors (KNN), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM) to evaluate the optimal combination of classification algorithm and sensor placement.

Sensor placement was associated with the classification performance (F = 7.39, p < 0.001), and the lower leg was the optimal sensor placement yielding the highest accuracy. However, neither the classification algorithm nor their interaction showed significant effects (p > 0.05). The best classification performance was achieved by the SVM using the left lower leg sensor with 90.1 % accuracy, 95.7 % sensitivity, and 84.1 % specificity.

Fall risk classification performance was affected by the placement of IMU sensors, with the lower leg showing the highest classification accuracy. Our results should provide insights to advance fall prevention technology in older adults.