Hi Eileen, thank you for your question. I have learned that motor complications of Parkinson’s, such as dyskinesia, can be quantified and identified from body-worn sensors. Within our sample population, the reliability of the system is similar to that of a clinician. This finding has a great implication for people with Parkinson’s: they can automatically have their symptoms monitored by daily wearing a watch-shaped sensor, without requiring personal monitoring from their clinician.

On a more personal note, I’ve really enjoyed getting to know the Parkinson’s volunteers in my studies! They are a highly motivated group of people, who strive for a better treatment or cure. Most volunteers are active in the community and are eager to give back, despite their impairments. I often find they challenge me to be a better person.

Kristopher, in our latest study, we are collecting periodic data that indicates off periods, mood, drowsiness, mental capability, and gait. We are interested in our ability to identify off periods, as well as fluctuating non-motor symptoms such as depression/mood and mental clarity, which greatly impact the quality of life in Parkinson’s. Clinicians treat nonmotor symptoms different than motor symptoms, so a mood tracking system would be valuable to them.

Regarding automatic medication dosing, probably the easiest way to do this is not through medication but through deep brain stimulation (DBS) settings such as voltage. DBS can induce dyskinesia when the voltage is too high, but the correct amount of stimulation can reduce tremor and other motor complications. The dyskinesia monitoring system could potentially act as a feedback control to adjust DBS voltage within an acceptable range.

Ayelet, Yes, it is possible to conduct a study that would quantify the efficacy of this system coupled with clinical treatment in reducing dyskinesia. Because the clinician is still an integral part of the treatment and would only use the system to inform them of patients’ symptoms, the study would need a control group that is treated by clinicians without access to information from the dyskinesia monitoring system. Also, because motor symptoms and complications fluctuate in Parkinson’s, the study would need to be conducted over several months or years to draw conclusions with greater certainty.

Regarding your second question, this is a really great question we have discussed at length with neurologists. Because each person displays different forms of dyskinesia, and variations also occur over the progression of the disease, the system would need to be “robust” in the sense of identifying all possible types of dyskinesia and classifying to the type that is being presented at the time of sensor observation. Our sample size of 19 participants is far too small to establish a power factor across the entire Parkinson’s population. Likewise, we have not conducted a study that tracks dyskinesia changes with time within individuals. However, the identification accuracy we reported on the poster is across all participants in the study, and is quite high. Preliminarily, we anticipate the system may be able to generalize across the entire Parkinson’s population if it was trained on a large and diverse enough sample population. One of the reasons we think this is because we are calculating a set of 20 features per sensor signal that are designed to capture the critical aspects of different types of dyskinesia reported in the literature. The machine learning algorithm has the ability to select the most important features in identifying dyskinesia across the sample population used to train the algorithm.

Timothy,
The decision tree classifier classifies 20 features per sensor (total of 100 features per participant wearing 5 sensors) into dyskinesia and nondyskinesia periods using visually observed dyskinesia periods as the target class. The decision tree trains on the data set, then validates using 10-fold cross validation. Each instance of 100 features are derived from 1 minute moving time windows of accelerometer data, so there is one instance per 1 minute window. All instances from all participants were included in the training and validation. For a nominal 1.5 hour observation period per participant, this totals about 102600 instances. The tree build nodes based on which feature reduced the classification entropy the most at the level of the node.

Excellent response! Thank you.

Mary, this is an insightful question. It is interesting because we are currently using observed dyskinesia as our “gold standard” to train our algorithms, even though inter- and intra- rater reliabilities are variable and well documented. When we use human-observed dyskinesia in the algorithm training, we would not expect our algorithms to have greater reliability than the human observers. You ask, is it possible to have greater greater reliability with the machine learning system? Yes, I suspect it is because the system is identifying quantifiable features (such a peak frequencies, energies, amplitudes, etc.) that are documented in the literature as describing dyskinesia, some features of which may not be detectable by human observers. To further investigate this possibility, we have plans to conduct cluster analysis, in which the machine learning algorithms groups features into clusters based on similarities in the data. Although the algorithm would not identify the clusters as belonging to dyskinesia or not, if we could correlate known dyskinesia periods to data within a cluster we could identify that cluster with dyskinesia. This method could potentially surpass the reliability of human observers.

Great idea – I’m already a MDS member, and have been collaborating with neurologists in developing the system. Broader input form neurologists would be insightful.