DISCLAIMER: This project is in no way associated with Microsoft Research. Furthermore, there is no rigorous scientific evidence beyond anecdote that this concept is scientifically repeatable or useful. The primary current goal of this project is to develop the hardware, software and procedure necessary to test the hypothesis that haptic feedback applied to the wrist of an individual with tremors can dampen/diminish some of the impacts of Parkinsonian/Essential tremors.

A variety of approaches have been explored to help manage the impacts of neurological tremors that impact up to 1% of people worldwide (including individuals with essential and parkinsonian tremor). Invasive methods as well as pharmacological solutions show promise, but can be cost-prohibitive and not easily accessible. Mild electrical stimulation to a users arm, coupled with real-time tremor monitoring with an IMU has proven to be successful in reducing tremor severity by roughly 40%, but this requires applying electrical signals that can be uncomfortable. An approach that is similar to this but replaces electrical shock with simpe, surface haptic motor vibration (i.e. putting a tiny motor on top of a users nerves) was developed in 2017 and is commonly known as PROJECT EMMA. This was a research project that garnered great interest from the Parkinson’s community. As recently as December of 2022 the only known prototype to exist is the Emma Watch. The publically available information that exists is in the form of press releases and media coverage which suggests that there are tiny, coin cell motors inside of a watch band. These motors can be pressed against an individual’s wrist when they are experiencing a tremor. Some references in media coverage to “machine-learning” suggest there is an algorithmic process to determine a vibrational frequency pattern for these coin cell motors (as controlled by an accompanying tablet) that mitigates impacts from tremors. It appears in videos and press that while the watch doesn’t completely negate a tremor, it does allow Emma, the recipient, the ability to draw straight lines. The conceptual framework believed to be operating is that the surface vibration of the coin cell motors on the wrists “short-circuits” the brain so that it doesn’t send as strong of a correctional signal. In principle, the vibrating watch concept is not dissimilar to the commercially available TouchPoints product, which uses wrist mounted units to apply low, medium and strong vibrations as a means to mitigate stress. While no rigorous academic studies confirm their efficacy for Parkinson's patients, case studies published on the web portal for the product indicate potentially positive impacts on users. As of the initial publication of this project page (5/17/2023) there have been no published followup studies into this technology. The only formal, published research study that has been conducted into this type of intervention found that it was not effective for individuals with essential tremor. They found no clear correlation between motor vibration frequencies applied to a users arm and tremor frequencies. This finding informs the design of this work in several key ways: 1) Parkinsons and Essential tremors may have different underlying mechanisms - perhaps something is unique in Parkinsonian tremor that makes surface haptics more effective 2) longer term vibration patterns yield different results than shorter term - a device that is simple to use for long periods of time will allow for gathering more data in real-life situations 3) tremor severity may be impacted by anxiety and conditions of testing within a lab environment - a device that is unobtrusive and usable in everyday life can reveal different (and I would argue more helpful!) results when compared to tightly controlled laboratory data - solutions in the lab are great to inform research but mean nothing if they don't apply to real-world situations 4) modern machine learning neural networks can reveal patterns in data that are unclear to other forms of analysis - machine-learning can fold proteins, fit regression problems and perform sophisticated classifications based on tremendous amounts of data. This project is an attempt to determine if the concept of the Emma Watch as described above has merit by creating the hardware and software workflow to allow anyone with broad maker skills to make and gather data.

Hardware:

The basic hardware is straightforward in principle and is well-suited to being created using hobbyist “maker” technologies. I provide links to all the items used for the initial prototype. I am constrained to purchase through Amazon by current ordering processes at my employer. Lower cost alternatives exist from different electronics vendors. In addition, a more customized solution using bespoke PCBs could optimize the design for both form and function. Such a project is unwarranted until this hobbysit proof of concept can be studied for efficacy.

  1. 10mm x 3mm 12000 RPM 3V Coin-cell vibrating motors - provides the vibrations on the users wrist
  2. DRV2605L coin-cell motor driver boards to control pulse frequency - allows for control of vibration patterns
  3. A potentiometer for manually scaling the strength of vibration - allows for manual sliding control of vibration patterns. A sliding control was chosen for ease of control when compared with a dial
  4. A LiPo battery - 3.7V, 500 mAh for powering motors and Nano
  5. Sparkfun Lipo Boost Converter/Charger - boost converter 
  6. Arduino Nano 33 BLE Sense Rev2 

SOFTWARE:

Using the Arduino to vibrate a motor is trivial. Using the Arduino to read accelerometer values is likewise trivial. The procedure for choosing an appropriate vibration pattern is less clear. This is where machine-learning can potentially assist. Recent advances in embedded machine learning allow for deploying advanced machine-learning models onto low-powered/low-cost processors. Rather than using a tablet for increased processing (as presumably done with the original Emma watch prototype), in this study, we will use the IMU on the Arduino Nano 33 BLE to collect kinematic data on tremor frequency using a 3-axis accelerometer that measures acceleration and gyroscopic motion of a user’s wrist in x, y and z directions. This data can then be analyzed in the frequency domain using signal processing with fourier analysis to quantify the frequency of the tremor, often using Power-Spectral Density methods. Various studies have correlated these (and other) different types of IMU data (both raw and heavily processed signals) with qualitative clinician assessments of Parkinsonian tremor severity (commonly the UDPRS scale). 

This project chooses to use a simple starting point making use of raw, unfiltered data for ease of implementation and to allow for more real-time processing of data. The Arduino Nano 33 BLE Sense is used in two software modes:

1) Data Collection Mode - simple collection of x, y, z values from the accelerometer and gyroscope obtained with and without varying vibration motor frequencies/patterns. Data from this will be used to train a bespoke machine learning model to optimize motor vibrations to minimize tremor impacts. This portion is functioning correctly and its implementation described in the procedure below.

2) Machine Learning Mode - the Edge Impulse Machine Learning model will be deployed to the Nano 33 BLE Sense to control motor output versus tremor input in real-time. This portion is currently in progress.

PROCEDURE:

This project will require the following from participants:

1) DATA COLLECTION: Participants will wear the wrist mounted IMU to gather baseline data on tremor amplitude and frequency using Edge Impulse. Participants will wear the IMU on their wrist while performing 1 of 5 classifications of activity in 10 second increments for a total of 2 minutes each. In total this will require 10 minutes of active time: 

  1.  idle, 
  2. moving their hand up and down, 
  3. moving their hand left and right, 
  4. lifting a glass
  5. lifting an eating utensil. 

Proof of concept training data along with a machine-learning model built to predict which of the above patterns is detected has been built and is available here.

Participants will then wear the wrist mounted IMU and have haptic feedback applied to their wrist while performing the same actions as above. The haptic frequency will initially be set to be equal to the dominant frequency of the individual’s unique tremor. Vibration frequencies will then be adjusted in reasonable increments to allow for development of a fitted machine-learning model to minimize tremor frequency based on input vibration. A randomized pattern of vibration will also be applied to compare with the non-random results. IMU data from this visit will be collected and fed into a Machine-Learning model developed using Edge Impulse to create a model linking motor vibration frequency to minimized tremor impact. This model will be deployed to the Arduino Nano 33 BLE Sense Rev2 for longer term use by participants.

2) MACHINE LEARNING DEPLOYMENT: The developed model will be deployed to the wrist mounted IMU device. The participant will be provided with the prototype watch and taught how to use it. The participant will use the device on their own over the span of a week to determine efficacy of the machine-learning model in minimizing tremor frequency from surface vibrating motors. Results will be analyzed and the process repeated/adapted as necessary. Hardware and software adaptations are likely to be needed and will be logged here.