Inspired by
personalized
medicine,
and as part of the
ADAPT in SC group,
we are researching the use of AI to train
individualized models for biomedical devices.
Devices empowered by this research learn to tailor monitoring and
treatment to the specific characteristics of an individual.
Training individual models
Classic methods for training AI models use group data.
It is cheaper and easier.
Collecting small amounts of data from each person
reduces individual burden while still providing
generalized performance across a population.
However, the model is now optimized for the group, not the individual.
Training an AI model on individual data
requires more effort in data collection because the
model must be retrained for every individual.
However, the model can be optimized to each individual, usually
outperforming the group model in predictive accuracy.
Group data
Individual #1 data
Individual #2 data
Case study: Detecting eating episodes by tracking wrist motion
In the eating
detection project
we use motion sensors in a smartwatch to track wrist
motion throughout the day.
There are 6 toal sensor axes, including 3 accelerometer (X, Y, Z)
and 3 gyroscope (yaw, pitch, roll).
Data is recorded at 15 Hz.
The below demo shows motion data being analyzed during a meal,
with moments of intake detected.
Motion sensors
Demo of wrist motion during a meal (click to play)
We train neural networks to detect wrist motions associated with eating
to identify timespans of those motions (e.g. meals/snacks).
The neural network analyzes a sliding 6 minute window of sensor data
and identifies the probability of eating P(e), where a higher value
indiciates a higher likelihood that the person is eating.
Periods of high P(e) can be detected as meals/snacks.
The below example shows a day in which the person self-reported 3
meals (green bars), which were all detected by our methods.
We also detected 3 false positives (orange bars A, B, C).
Neural network analyzes wrist motion
Example detection of periods of eating
(green = true positive, orange = false positive)
Data
We have 2 datasets that can be used to train eating detection models.
The Clemson
all-day dataset (CAD) contains data for 351 people wearing a wrist motion
tracker for one day each.
The Clemson
multiday dataset (CMD) contains data for 9 people, each wearing
a wrist motion tracker for 7-21 days.
We can use CAD to train a single group model, and CMD to train
individual models for each person.
Code
Code for training our models is available
here.
It uses TensorFlow and
is programmed in Python.
It includes functions for loading the data, preprocessing,
model training, and inferencing for testing.
It also includes functions for calculating metrics to evaluate
model performance.
Results
The individual models perform better on average because they can
learn motion patterns that each individual uses for eating,
for example actions that may be associated with utensil and food
preferences.
More importantly, the indivudal models also learn what motions
are NOT associated with eating, in other words what sorts of
activities the individual engages in throughout the day.
These types of motions are extremely varied across a large group
of people, but can be significantly narrower for each individual.
This decrease in variability can help reduce false positives.
Weighted accuracy (20:1 to account for class imbalance) vs detection threshold Ts
