NeoPET

The Novel Application to Parkinsons and Essential Tremor (NeoPET) provides an innovative, seamlessly-integrated way of tracking Parkinson’s and Essential Tremor symptoms and treatment responses through a smartphone. By utilizing machine learning to analyze and quantify symptom severity, NeoPET is able to provide critical, time-sensitive information to contextualize disease progression and treatment response. The portal offered by NeoPET allows physicians to be alerted to urgent events and evaluate treatment options at any time

Inspiration

In the U.S., each year, 60,000 Parkinson’s cases are added to the 1,000,000 currently diagnosed patients, while only 16,000 motor specialists or translational DBS clinicians, often consolidated around urban areas, are available to treat them. [1] These temporal, geographical, and physical limitations provide an extenuating frustrating experience for many PD and ET patients alike, especially for first visits or if they feel a treatment is not working — or worse. Oftentimes, these specialists are booked months in advance for new or existing patients, making both prompt and accessible intervention extremely difficult. [2] Parkinson's Disease (PD) and Essential Tremor are dynamic in both how they present themselves in patients and how the research community quantifies severity and progression, meaning existing data that cannot cross platforms is often lost. Furthermore, treatment options such as L-dopa or deep brain stimulation (DBS) settings can have rapid, drastic effects that do not set in until hours, days, or weeks after implementation. We identified a need for an app that would provide physicians with the real-time data needed to modulate treatment plans and intervene on rapid progression without the need for an in-person visit.

What it does

The app passively and actively tracks various features that are used to evaluate the severity of Parkinson’s in clinic through day to day movement and activities. Primary complaints from PD and ET remote tracking methods are constant device interactivity and physical discomfort of a wearable, so we strive to combine a pleasant, almost completely background process for the user while not sacrificing accuracy and reliability. Phone sensors can flag trigger events with strong confidence to begin or terminate a certain datastream collection. The amalgamation and detailed breakdown of these results posted at the periodicity of collection can be viewed through a physician portal. This implementation enables physicians to continuously monitor their patients post treatment with little to no inconvenience for the patient and provides the clinician with more data to drive better clinical outcomes and more efficiently optimize treatment plans.

How I built it

Using Android Studio, we developed an app that implementes the capabilities of smart phones to serve as a diagnostic tool for symptom severity. By passively tracking the user’s typing habits and gait through smartphone keyEvents, accelerometers, proximity sensors, quaternions, and gyroscopes, we obtained information which would be automatically uploaded onto the cloud. When our sensor pipeline, running as a background service, senses phone-in-pocket and several steps, recording of the IMU-6/DOF initiates, and terminates after a 5 second pause in motion or 2 minutes of streaming. This gait data is then processed in MATLAB, then extracting 20 time features and 9 frequency domain features, all significant to the progression and state of PD and ET. This includes Stride Time, Stride Variance, omnidirectional Amplitude Variance, omnidirectional peak frequency, and omnidirectional median frequency. These are both raw classifiers researchers use and also features which can be imported into an SVM if given more grouped data.These were not able to be detailed in the video due to time constraints but we are happy to answer questions regarding their use and relevancy. For a more comprehensive patient profile, we also combined two too-often issued features into a symbiotic stream. We offer a patient journal and mental state survey in the local app, recommended to be taken once a week or more often. However, PD adherence to tasks is low, especially surveys as most people feel they are not meaningful, even if they are delivered standardized as instructed by UPDRS. To combat this and collect data efficiently, built into this app we have an objective bradykinesia and tremor+handedness analysis using keystroke events to classify 9 hold-time and 18 latency features. Supporting this leveraging function is a database with PD and Naive keystroke data, which MATLAB then used to normalize the data, reduce the dimensionality using a Linear Discriminant Analysis technique with pseudolinear discrimination, and train several binary classifier support vector machines on data subsets using a standardized gaussian kernel and radial basis function, which yields 98.08% accuracy for overall PD presence, 94.12% accuracy for handedness of bradykinesia and 96.37% accuracy for tremor presence. This is following a k-fold cross validation (k = 10) to ensure over fitting is not happening. These data streams are then stored and viewable in a classified and succinct way for the clinician to skim or deeply analyze themselves.

Challenges I ran into

The upload and download from data servers to move the data between the app, computer, and website proved difficult because MATLAB did not have native functionality for Google Firebase. Developing an optimal LDA and SVM was an issue, and we had to make do with the data out there, meaning not all gathered functions could have percentile features (percentile severity is a growingly preferred method of continuous progression over discrete) based on SVM’s. Had we more time or perhaps data specifically, we'd have attempted to use accelerometer data for hand tremors as well, but databases are not accessible without requesting from the owner or author, which is not within the timescope of MedHacks.

Accomplishments that I'm proud of

We are proud of our resilience. Many times we did feel as though we’d gone over our heads, but we persevered past these doubts and misgivings and created a product that we are proud of and believe can make a difference in the lives of Parkinson's patients. We take pride in our ability to rise to the challenge and develop a functional app with multi-source data acquisition capabilities within 48 hours, having little prior experience.

What I learned

Android app development and how to utilize web APIs for data transfer.

What's next for NeoPET

In the future, we would like to implement more functionality to our app such as encrypted call data processing to assess dysarthria, but privacy concerns pushed this out of the scope for now. We consider in the future, if we gain traction, implementing a voice, emotion, and facial masking analyzer (other important parameters) that works in the background of virtual interactions with clinicians for telehealth appointments. We also found ANNs can be used on accelerometer data to classify the likely “location” (bed,pocket,bag,etc) for the phone, which would provide an even more accurate trigger event system. Additional features would improve our classifier and provide more valuable feedback to the physician.

Works Cited Statistics. (n.d.). Retrieved September 06, 2020, from https://www.parkinson.org/Understanding-Parkinsons/Statistics

Chard, G. (2014, May 1). Telehealth to Digital Medicine: How 21st Century Technology Can Benefit Patients (Issue brief). Retrieved September 6, 2020, from https://docs.house.gov/meetings/IF/IF14/20140501/102173/HHRG-113-IF14-Wstate-ChardG-20140501.pdf