Automating the Clinical Assessment of Independent Wheelchair Sitting Pivot Transfer Techniques

Using proper transfer technique can help to reduce forces and prevent secondary injuries. However, current assessment tools rely on the ability to subjectively identify harmful movement patterns.

The purpose of the study was to determine the accuracy of using a low-cost markerless motion capture camera and machine learning methods to evaluate the quality of independent wheelchair sitting pivot transfers. We hypothesized that the algorithms would be able to discern proper (low risk) and improper (high risk) wheelchair transfer techniques in accordance with component items on the Transfer Assessment Instrument (TAI).

Transfer motions of 91 full-time wheelchair users were recorded and used to develop machine learning classifiers that could be used to discern proper from improper technique. The data were labeled using the TAI item scores. Eleven out of 18 TAI items were evaluated by the classifiers. Motion variables from the Kinect were inputted as the features. Random forests and k-nearest neighbors algorithms were chosen as the classifiers. Eighty percent of the data were used for model training and hyperparameter turning. The validation process was performed using 20% of the data as the test set.

The area under the receiver operating characteristic curve of the test set for each item was over 0.79. After adjusting the decision threshold, the precisions of the models were over 0.87, and the model accuracies were over 71%.

The results show promise for the objective assessment of the transfer technique using a low cost camera and machine learning classifiers.