Combining a commercial shoe based wearable system and a simple machine learning model to classify activities of daily living

INTRODUCTION
Osteoarthritis (OA) is the most prevalent chronic joint
disease affecting 250 million people worldwide [1]. The
disease is characterised by the degeneration of
articular cartilage and can often be accompanied by
pain. The duration and frequency of patient’s clinical
assessments are being increasingly limited by
financial and infrastructure-based barriers.
Wearable sensors systems, such as pressure insoles,
could be used to assess patients habitual loading
patterns throughout their daily routine. The first step
towards having the full picture of daily skeletal loading
is knowing how much time is spent completing
different movements that load the joints in the lower
limb.
Therefore, the aim of this study was to evaluate
whether a minimally obtrusive and commercially
priced insole system could be combined with a simple
machine learning model to identify typical activities of
daily living.
METHODS
Fifteen healthy participants (six females) were
recruited to participate in the study (age: 42 ± 16
years, mass: 77.6 ± 16.7 kg, height: 1.78 ± 0.11 m).
NURVV Run, a commercially available fitness monitor
system integrating thin pressure insoles (each
embedding 16 force sensitive resistors) and a IMU
were fit into the participant’s daily shoes. The data
collection protocol consisted of three phases. Firstly,
the participant completed level ground walking around
the lab space. Secondly, participants were instructed
to stand from a seated position, pause for a few
seconds before returning to a seated position and
repeating. The final movement was ascending a ten-step staircase, pausing at the top before descending
the same staircase and repeating. Each of the three
phases was repeated for two minutes to collect several
examples of each movement. The dataset was first
segmented into the five activities of daily living:
walking, standing, sitting, ascending stairs, and
descending stairs. The initial contact and toe off time
points were identified through a thresholding approach
based on body weight. Then contact periods were
extracted and normalised to 101 data points.
A simple three-layer neural network was created using
the package Pytorch within Python to perform the
supervised classification task. Rectified Linear Unit
(ReLU) activation functions added non-linearity to the
output of the first two layers before a probability of the
input data belonging to each of the five classes was
output by the final layer. A total of 21 time series
signals were input into the neural network model: the
16 individual pressure channels; the total pressure
from all channels, the anteroposterior and
mediolateral position of the centre of pressure (CoP),
and the anteroposterior and mediolateral distance that
the CoP had moved since initial contact. Prior to input
into the model each feature is independently
standardised to z-scores using the training means and
standard deviations. The model was trained and
evaluated using a leave one subject out approach,
where all but one participant’s data makes up the
training set, the trained model is then evaluated on the
remaining participant’s data left in the validation set.
This process is then repeated for all participants and
accuracy is averaged across each participant.
RESULTS & DISCUSSION
The average classification accuracy across all five
movements was 92.5%. The standing and sitting
movements were the most accurately identified, with
slightly lower classification for the three forms of
ambulation. Walking was occasionally mislabelled as
stair ascending (10.4%) or descending (4.5%), as
might be expected due to the possible similarities of
the CoP movement between the three movements.
Figure 1: A confusion matrix showing the
classification performance for each of the 5 activities.
CONCLUSION
This study showcases that pressure data from a
commercially priced wearable system could be used
to accurately segment data from daily living into key
joint loading tasks. Knowledge of the time spent
completing each task throughout a day could be paired
with load predictions, driven from the pressure data, to
estimate patient load accumulation more accurately.
REFERENCES
[1]. Prieto-Alhambra, D., et al. Annals of the
rheumatic diseases, 73, 1659-1664, 2014. [LIN