Demonstration of model prediction
Here we showcase a custom-made convolutional neural network (CNN) model's predictions for primitive action identification in stroke rehabilitation. This robust model is designed to compute separate embeddings of different physical quantities in the sensor data, coupled with an innovative use of instance normalization (over batch normalization).
This approach significantly enhances the model's adaptability to new patients, demonstrating resilience to possible distribution shifts. Remarkably, the CNN model has achieved an average classification accuracy of 70%, which is a substantial improvement over existing methods.
For this demo, we display the model's predictions in action. Moreover, to ensure even greater accuracy, we employ a post-processing technique using a sliding window for smoothing, which refines these predictions to more accurately represent the rehabilitation activities performed by the patients.