Finger taps – large data sets required to enable machine learning.

Published by:
Published on:
Published in:

Koen Van den Brande CSO

Parkinson’s is diagnosed and progress monitored by means of ‘clinical observations’ …

Occasionally, as in the run-up to DBS, those observations are recorded on video.

Most of the time they are not, in my experience at least.

The UPDRS – the universal rating scale is used by medical professionals and during drug trials.

In a speech at the 2019 WPC in Kyoto, Professor Bas Bloem recalled how, during his training as a neurologist – at one of the best centres in Europe – his professors would routinely disagree in practice about one of these ratings – the ‘finger tap’ …

The reason is easy to see …

UPDRS instructions on ‘finger tapping’

Try scoring a typical observation of yourself …

And yet this score is critical for the assessment of a PD patient.

Now, we know that if we could use machine learning, we would probably be able to massively improve the accuracy of these ratings.

However, for machine learning to be effective we need large data sets.

In its simplest form we would need a lot of finger tap recordings of PD patients and a control group – maybe the patients’ carers. That would enable the AI engine to be taught the difference between a finger tap of a PwP (Person with Parkinson’s) and a person who is PD-free.

The next step would require some expertise and manual effort.

A group of neurologists would have to review a sufficiently large subset of the observations and record consensus ratings.

Based on that, the AI engine could then be taught how to rate ‘finger taps’ accurately from video recordings.

So we have decided to try and put this into practice in as ‘low-tech’ a manner as possible.

We are using smartphone cameras to record observations and send them to two WhatsApp groups – one for PD Patients the other for PD Carers, using their personal smart phone devices.

That results in a collection of video clips of a few seconds, with date and time stamp.

At the moment we are not yet trying to link observations to Patient data on medication, symptoms  etc. That will require more of the PLM infrastructure to be put in place.

We may be able to use a low-tech approach to this too, but I propose to first try this out in the context of a hospital with a PD research team in place. There we can then provide formal training on how to make the recordings, ideal timing and other details to be added.

This will be the ‘rollout phase’, as we start to systematically recruit large numbers of patients, as they visit their neurologist.

For now we are simply using a group of volunteers keen to make a start with proving this can be done.

We will use the following guidelines

  1. Get someone else to operate the camera.
  2. Attach the recording directly in a WhatsApp message in the PLM trial PD or Carer group
  3. Record each hand separately
  4. Use a featureless background as much as possible

By taking part, the individual consents to this data being used for research purposes.

Candidates can join the relevant WhatsApp group by sending me a contact message.

I look forward taking this on …

European Headquarters

Belgium:
NeuroPath
La Lanterne, Rue Jean Burgers 2, 7850 Enghien
Technologielaan 9, 3001 Heverlee
Email: contact@neuropath.life

North American Operations

United States:
NeuroPath US, Inc.
222 W. Merchandise Mart Plaza, suite 1230
Chicago, IL 60654
Email: contact@neuropath.life