Instrumentation & Measurement Magazine 26-4 - 42

posture but with several missing detections and reaches ADRs
of 66.67, 82.35, 82.35, 94.74, 91.89, respectively. However, the
ICA-KD-based scheme provides better performance since it
precisely identifies this improper sitting posture with a good
detection rate. This scenario confirms the superior performance
of the ICA-KD scheme compared to the other methods.
Scenario 3: Sitting Posture with Forward Pressure: Finally, we
consider the scenario when the improper sitting posture is due
to applying high forward pressure. As summarized in Table 1,
the ICA-KD approach achieved better detection results with
a high detection rate of 98.25%. Visually, we can also see that
ICA-KD reaches superior performance compared to the other
models (Fig. 4). The PCA-based conventional indicators demonstrate
poor detection performance as they fail to detect this
abnormal posture. Here, the red shaded zone highlights the
period of the abnormal posture. We omit the visualization of
the results of the other sitting postures because they all provide
relatively similar conclusions.
The ICA-KD approach can detect the presence of abnormal
sitting positions but does not identify the types of the
detected sitting position. Identifying the type of sitting position
is accomplished by analyzing the contribution plots of
each variable. Fig. 5 demonstrates contribution plots for improper
sitting postures in the wheelchair. In Fig. 5a, we observe
a larger pressure level in S1, which is placed under the right
leg, than the pressure levels in S2 and S3, which indicates the
presence of a wrong sitting posture inclined on the right side.
Similarly, for sitting position 3, we observe a larger pressure
recorded by the sensor S3, which is placed under the left leg,
compared to the pressure levels recorded by the other sensors
(Fig. 5b). In a similar way, from Fig. 5c we observe higher forward
pressure in S1, which is placed under the coccyx, than
pressure levels recorded by S1 and S3. Importantly, we can see
that using the ICA-KD approach and the contribution plot, the
proposed strategy can detect and distinguish the type of sitting
posture for wheelchair users with unlabeled data.
Conclusion
The number of people using wheelchairs has increased in the
last few years, and hence, identifying improper sitting postures
in wheelchairs is necessary to avoid health issues. This
paper proposes an unsupervised monitoring approach that
integrates the ICA model with a KD-based abnormal event
indicator to monitor improper sitting postures in a wheelchair.
The proposed ICA-KD method used the normal event
data from pressure sensors embedded in the wheelchair to
develop an ICA model, which was then used for identifying
the incorrect sitting postures using KD statistics. The ICAKD
approach over-performed the conventionally applied
abnormal event indicators of ICA and PCA-based methods
in detecting improper sitting postures. The ICA-KD strategy
reached high detection with an F1-score of 99.41%, better
than the conventional anomaly indicators. Hence, integrating
the ICA model and KD-based detector is a promising tool
for detecting abnormal sitting postures in a wheelchair. Furthermore,
a contribution plot was conducted to discriminate
between distinct sitting postures. In short, the proposed ICAKD
approach can detect and distinguish sitting postures for
wheelchair users.
Despite the enhanced detection performance achieved by
the proposed approach, future works will improve its ability
to automatically discriminate different abnormal sitting types.
One possible way to alleviate this limitation is to incorporate a
classification stage based on a support vector machine or other
classifier applied to detected sequences. Another alternative is
to apply univariate monitoring charts, such as the generalized
likelihood ratio test [11], to the residuals from the ICA model.
Furthermore, another direction of improvement consists of
using data augmentation techniques to generate large-sized
data, which improves the construction of models and thus enhances
the detection process.
Furthermore, it is interesting to incorporate more information
from posture variations, such as the use of armrests, and
foot position, to further improve abnormal postures recognition.
Also, other types of sitting postures could be considered.
Nowadays, numerous small sensors are available and can
be embedded in daily gadgets, like smartphones and smartwatches,
which enable the collection of different variables to
monitor the state of wheelchair users. Future work will investigate
including more data inputs, such as heart rate and blood
pressure provided by a smartwatch or a smartphone, for further
improving the efficiency of sitting posture recognition.
Fig. 5. Variables contribution plot: (a) 2nd position at sample point 100; (b) 3rd
position at sample point 200; and (c) 4th position at sample point 250.
42
Acknowledgments
This work is supported by King Abdullah University of Science
and Technology (KAUST), Office of Sponsored Research
(OSR) under Award No: OSR-2019-CRG7-3800. The authors
would like to thank the Manipal Institute of Technology of the
Manipal Academy of Higher Education for supporting this
work.
IEEE Instrumentation & Measurement Magazine
June 2023
Instrumentation & Measurement Magazine 26-4

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 26-4
