_edited.png){kind=small}
RESULTS
Our algorithm successfully serves as a PT assistant for pallof presses with a classification accuracy of 87.7%. We tested our algorithm with data for a correct, incorrect with error in the x-direction, and incorrect with error in the z-direction pallof press. Additionally, this data was collected using a variety of resistance band levels and with different people to create a more representative data set. This model serves as a proof of concept that PT Assist is a viable product that can help patients improve their physical therapy outcomes.
​
Fig. 11 shows a confusion matrix of our results. As you can see, the top right to bottom left diagonal has the highest numbers, showing our algorithm has a high accuracy. Our most common source of misclassification is from our algorithm classifying the trial as incorrect from not being perpendicular to the body (error in the x-direction) while it was actually a correct trial of the pallof press. We expect that with more data collected, we could calculate a better RMS cut-off value, leading to more accurate classifications.
Fig. 11. Results confusion matrix.
ERRORS AND NEXT STEPS
1
Currently, our algorithm only identifies the two most common errors patients encounter when performing a pallof press. We hope to consult more physical therapists and collect more data to expand our analysis and algorithm to identify and correct more sources of error.
2
To increase the accuracy of our classification, we want to develop and compare different classification algorithms. These include testing different methods for determining the cut-off frequency, and using eigenvectors and principle component analysis to create a basic machine learning model.
3
After these steps are complete, we want to expand PT Assist to additional exercises. After consulting with physical therapists about the most common exercises used, we would follow a similar process as we did for the pallof press. This would increase the value to the users and make PT Assist more versatile.