Lauren & Mick Questions Flashcards
1) Why is the emphasis on movement quality so important?
If movements are undertaken using too much compensation that is not recommended, there are long-term implications such as: pain and joint contracture. Getting high dose is recommended in stroke rehab but there is a suggestion that moving patients towards more “normal” pre stroke patterns of movement is good.
2) If you had limitless resources what would you do to get sonic sleeve to the point where it could be implemented in the home at scale. What features would it need to have etc…
- Make the system adaptive - increase or decrease based on the success or failure of patient to stay within thresholds that are set - this means intensity can be set at a “just right challenge”
- Run the system off mobile phones (as other companies are doing for healthy participant tracking - such as yoga postures etc). Posenet from Google using machine learning in the web browser, is a good candidate to build Sonic Sleeve and extend into the home of anyone with a smartphone.
- The ability for therapists to monitor patients in the home via dashboards or similar - if all data can be tracked, then this can provide financial rewards via codes for claiming on health insurance etc and could help to encourage more use
- Definitely permit Spotify or YouTube music playlist integration to ensure interest is retained for the longer term. Also, this will create another feature which is variation in speed.
- Variation of movement type - train models on reaching on more levels than 1 as the system currently tracks.
Extend to track lower body for sit to stand etc as Upper Limb though important, is not the full rehab patients require.
3) What you’ve done with the sound, binary on off is pretty basic, what other aspects of sound manipulation have been used successfully in the literature?
Self-chosen music linked to higher levels of motivation ->
Page: 22 Music Supported therapy (MST) fine and gross motor movements - playing drums and pianos - (Schnieder et al., 2007). Harnesses motor learning principles. - high dose. Evidence that music may play a role in enhanced rehabilitation from Tong et al where a muted group did not see as much benefit as the music group. (Tong et al., 2015). Maybe it is motivation only and the mute group were just not engaged in the exercises?
Page 24: Enhanced cortex activation in MST - Amengual et al., 2013
4) Where do you think this field is heading based on what you know of the literature? What are the main opportunities and future directions?
There are two key areas: the technology using machine learning or other algorithmic approaches to monitor and track in real time a patient’s progress and embedding real-time feedabckj into the systems to help guide movement quality, reps and duration.
Monitoring patient progress in the home where most rehab takes place is very important and the natural way the field is going - low cost tech into the home Saving time for patients and therapists while retaining high-quality rehab. A focus on movement quality in the home is important, and most companies in this area have not achieved much in this area. The RATE paper by Edward Averell and yourself Frederike looks like an interesting approach using algorithms to detect compensatory movements.
Music has been shown to be a versatile stimulus that can tap into the reward system of the brain - Blood and Zatorre 2001 paper highlighted this. Integrating more machine learning models with personalisation - targetted playlists and remote therapy monitoring are areas of development in the US in particular. Permitting therapists to claim back costs for the ability to track patient progress over time. Music could help encourage better adherence to rehab and allow therapists to push the intensity, dose and movement quality up for patients.
5) Given that interactive machine learning is about building systems that non-experts can use to train their own machine learning models to identify and solve their own problems, what can you say about how your project adds to this body of knowledge or otherwise?
I was able to collaborate directly with staff and patients who had no experience in machine learning whatsoever. I think the project provides support for IML approaches to test and build new systems that can be used clinically. Interactive machine learning permits the co-design of systems, and this current research showcases one model - working with end users, clinicians and designing and testing models on the fly helps to rapidly change and update a system.
The data collected involved the use of Wekinator and Interact ML. The system that went into the homes of patients was designed to be used by complete novices.
However, for therapists to train the system would involve further GUI design and testing. And further optimisation of the platform.
6) Is there any specific IML knowledge that your project has contributed that you think might benefit other researchers using IML in the context of clinical work specifically?
- can use an iterative approach to system design using machine learning live with clinical experts - for example, in the workshops where I could update the models on the fly shows that novices can directly influence a system as it is being built. placing the users (therapists and patients into that process).
- From training models directly with staff on the stroke unit, it is clear that the technology needs to be accepted and adopted by the therapists for the technology to get used. Their time is very constricted so having the ability to use models that train quickly is good - but this will not scale well. Longer term, a system such as Sonic Sleeve would need templates that a patient can calibrate with minimal staff input. Setting thresholds requires the staff and it is important to keep therapist reasoning in the loop so this current research has really highlighted just how challenging it is to embed technology into a stroke unit - or into the home environment.
7) Can you describe in more detail how IML was considered an approach as part of your experimental design and the extent to which IML related outcomes were considered as part of your analysis?
- When the project was being setup it became clear we needed to have the ability to create individualised models fo movement for each patient. So interactive machine learning was a natural choice using Wekinator permitting fast training directly on the stroke unit at Queen Square.
- The analysis was more proximal to the thresholds that were set by health experts - which gave the primary outcome variable for analysis. The point of the research was not to validate one ML model or approach over the other but to make use of models that could be used quickly to provide feedback in real time to patients and provide a proof of concept using interactive machine learning with supervised learning.
8) Are there any results that might be of interest to psychology researchers who are also using IML in clinical settings and that you might have been able to expand upon?
Using the data collected as a training set for more generic clinical models that could permit scaling up the system would be an interesting extension. Other researchers may want to see if a patient could be categorised using the data to set thresholds for feedback for example.
9) What challenges to IML in general did you face (e.g. from participants, collaborators, users)? Did you manage to overcome these? Was it basically impossible for practitioners or patients to use your system without you being present? Was this important to you at all? If not, why not?
- Page 122: It was a challenge to track trunk flexion using the lab based OpenPose system - the camera had to be positioned at 2 oclock for right side tracking and 10 o’clock for left. This means bimanual tracking would not work well with only one camera.
- Page 122 The at home version had some issues with Windows 10 crashing at times
- Page 99: Limitation of having to spend time calibration and setting thresholds to each individual patient - this can’t scale up well.
- There could be an issue of “overfitting the data” when translating to other users based on the data collected.
However, in the context of the study having data fitted to the individual to create realtime feedback seems to make sense.
It was a real challenge to train the system, and the current system for both lab and at home versions required a researcher to help setup and train the models. However, once the models were trained the at home version was suitable for a patient to launch and train without anyone supporting.
- It was not possible for the system to be trained by therapists and patients without support in this current iterations - however a GUI could be created in future to permit this and the Unity project was getting close to permitting a novice user to train and set thresholds for the models.
- However, there were issues -for example if a patient moved very slowly in their movements in the at home system - the amount of data collected started to strain the system - in these cases, deleting and collecting less data was a solution. It could be possible to downsample the amount of data in future iterations for those with severe impairments.
10) Where in the thesis do you address challenges when using IML?
- Page 122: It was a challenge to track trunk flexion using the lab based OpenPose system - the camera had to be positioned at 2 oclock for right side tracking and 10 o’clock for left. This means bimanual tracking would not work well with only one camera.
- Page 122 The at home version had some issues with Windows 10 crashing at times
- Page 99: Limitation of having to spend time calibration and setting thresholds to each individual patient - this can’t scale up well.
- Page 52: Retraining often required - this meant Wekinator was a good choice as retraining was generally quite quick. The at-home version using Interact ML was also quite quick - although the models needed to be retrained even more
