PhysiQ: Off-site Quality Assessment of Exercise in Physical Therapy
Co-author: Meiyi Ma
Welcome to a new world of digital physical therapy — PhysiQ. As a groundbreaking technology, we’re set to revolutionize physical therapy by ensuring patients can effectively continue their therapy at home. I’m excited to introduce our new framework that addresses a critical challenge in the world of physical therapy — off-site quality assessment of exercise.
When you think about physical therapy, you probably imagine a patient in a rehab center, diligently performing exercises under the watchful eye of a therapist. This supervision is critical as it helps ensure the correct execution of exercises, which in turn leads to optimal recovery. But what happens when the patient goes home? Are the exercises performed accurately without expert supervision? The truth is, for many, the lack of supervision, quality assessment, and self-correction lead to inaccuracies in posture and performance during at-home exercises. This is a significant issue, considering that patients spend the majority of their recovery time outside the clinic.
Despite the advances in technology, there hasn’t been a comprehensive solution to this problem. Indeed, we have Human Activity Recognition (HAR) in wearable devices that can recognize basic day-to-day activities. However, these technologies don’t cater to the specific needs of therapeutic rehabilitation. There have been some attempts at exercise quality tracking using vision-based devices, but these are cumbersome to set up and not user-friendly, especially for those with limited mobility.
This is where PhysiQ steps in. Our new framework uses passive sensory detection to track and quantitatively measure off-site exercise activity. We’ve employed a novel multi-task spatiotemporal Siamese Neural Network that evaluates exercises based on a person’s absolute and relative quality of exercise performance. With PhysiQ, we aim to provide a much-needed solution that will help patients effectively monitor their at-home physical therapy exercises, ensuring their rehabilitation process continues smoothly outside the clinic. Again, we are not trying to replace physical therapists but to enhance the experience of PT outside of the clinic. :)
In our research, we’ve collected and annotated motion data from 31 participants with different levels of quality, ensuring our model recognizes the nuances in exercises, works with different numbers of repetitions and is adaptable to individual PT progress. Our evaluation results have been encouraging — PhysiQ has achieved an accuracy of 89.67% in detecting levels of exercise quality and an average R-squared correlation of 0.949 in similarity comparison.
But, as with any groundbreaking innovation, there are challenges ahead. We’re still exploring how to digitize physical metrics, how to tailor the model to the individual differences among patients, and how to collect sufficient annotated data. Nevertheless, we’re excited about the potential of PhysiQ, and we look forward to sharing our progress as we continue to innovate in the world of physical therapy.
Our first and foremost achievement is the development of PhysiQ, the first-of-its-kind framework for the quantitative measurement of exercises using a smartwatch. PhysiQ stands out in its ability to identify and digitalize three key exercise metrics: range of motion, stability, and repetition. These metrics in exercises provide a detailed understanding of the functionality of the skeletal system, enabling patients to effectively monitor their at-home physical therapy exercises.
Central to PhysiQ is our multi-task spatiotemporal Siamese Neural Network. This innovative neural network measures both the absolute quality and relative quality of an individual’s physical therapy progress. It paves the way for patients to understand the quality of their off-site exercises over time, equipping them with valuable feedback that can significantly enhance their recovery process.
We have also developed an application that collects users’ motion data via a smartwatch and provides real-time, explainable feedback. This feedback is based on their quality of exercise and comes with recommendations to help them improve their performance.
To ensure our model’s robustness and reliability, we have collected and annotated motion data from 31 participants, who performed three shoulder exercises (shoulder abduction, external rotation, and forward flexion, as depicted below) with different levels of stability, range of motion, and repetition.
Finally, we conducted an extensive evaluation using real user data. The results have been impressive, with our framework outperforming the baselines by 47.67% on average in R-Squared for all exercises and all three metrics.
We also conducted a user experience study to understand how user behaviors influence the framework. The insights gained from this study will help us refine and improve PhysiQ in the future.
Let’s delve a little deeper into how PhysiQ works, as shown above.
One of the critical components of our system is the multi-task spatiotemporal Siamese Neural Network (SNN). In this section, I will give you a sneak peek into its structure and functioning, giving you a clearer understanding of what makes PhysiQ so unique.
Let’s consider two exercise instances, referred to as the ‘signal’ and ‘anchor’ exercises. These two one-repetition exercises are fed into our network. The signal exercise can be compared against the anchor and vice versa, providing a dynamic basis for analysis.
Next, we utilize a technique known as Sliding Windows Segmentation to break down the continuous exercise data into manageable chunks. This is followed by a process of Spatial and Temporal encoding, which transforms these segments into a format that our model can understand and process.
These encoded exercises are then fed into an attention mechanism. This mechanism uses learned weights to focus on the most relevant features of the exercise data. The output of the attention mechanism is a high-dimensional representation of each exercise that encapsulates its most critical aspects.
Now, it’s time to determine how similar these two exercises are. We compare these two high-dimensional representations and output their similarity score using cosine similarity. This score provides an indication of the relative quality of the exercise based on an individual’s PT progress.
But we’re not done yet. There’s another critical aspect of the framework that focuses on absolute quality. The high-dimensional representation of the signal exercise is also processed into a Multi-Layer Perceptron (MLP) network. The MLP network’s output is a classification result based on the three key metrics: range of motion, stability, and repetition. This gives us a quantitative measure of the absolute quality of the exercise.
By combining the results from these different components of our SNN, we provide an overall evaluation of the quality of off-site exercises. PhysiQ’s strength lies in its ability to deliver an evaluation that is both comprehensive and specific, addressing the unique needs of individuals undergoing physical therapy.
In conclusion, PhysiQ heralds a significant advancement in the realm of physical therapy. By enabling patients to take charge of their recovery process, it offers a level of empowerment that has been missing in traditional rehabilitation programs. Moreover, it provides invaluable data to therapists, enabling them to personalize each patient’s therapy plan for optimal results.
We are relentlessly refining and improving PhysiQ, guided by user feedback, research findings, and technological advancements. As we move forward, we are tremendously excited about PhysiQ’s potential to revolutionize physical therapy, ensuring that patients get the most out of their recovery journey.
If you’re intrigued by PhysiQ and would like to delve deeper into our research, you can read our complete paper here. It provides a comprehensive look at our innovative framework, along with detailed insights into our methods, evaluations, findings, and aspirations for the future. We welcome your thoughts, questions, and feedback as we continue on this exciting journey of technological innovation in physical therapy.
Lastly, thank you for reading my blog!
