Page 145 - AIH-1-2

Artificial Intelligence in Health Movement detection with sensors and AI

For example, a value of 10 in the off-diagonal position with appropriate tools. One such limitation concerns
indicates that 10 instances of “Breathing” were incorrectly the utilization of duct tape to secure the monitor to the
predicted as “Seizure” by the classifier. The implications of mannequin’s chest. Since duct tape is impractical for
these findings from the confusion matrix could be twofold: live patients, an alternative solution, such as a strap-like
Classes 0 and 1 are well-recognized by the model, possibly device, could be introduced to ensure the monitor remains
because these movements possess distinct characteristics securely in place on the patient’s chest without causing
that are easily discernible by the sensors. Conversely, classes discomfort. Another limitation pertains to the battery life
4 and 5 have the lowest instances of correct prediction of the sensor device. Throughout the study, it was observed
(17 and 20, respectively), suggesting that recognizing that the sensor’s battery depleted to 50% capacity after
“Breathing” and “Seizure” movements presents greater approximately 3 h of use. This limitation could be resolved
difficulty for the model. This challenge could stem from by having the patient’s nursing team replace the sensor at
similarities in the patterns of these movements or subtle the 6-h mark. The replaced sensor can then be recharged
characteristics that were not captured by the sensors or and reused for subsequent monitoring sessions.
model features. Misclassifications (off-diagonal numbers) Inpatient falls pose a potential risk of prolonged hospital
highlight areas where the model confuses one movement stays, further injuries, and various complications for many
for another. Although all off-diagonal numbers are 10 or hospitalized patients in recovery. Hospital-onset seizures
lower, indicating a decent level of precision, these errors also present a concern, as they are difficult to predict and
could carry significant consequences depending on the can significantly worsen a patient’s overall well-being.
clinical importance of the movements. For instance, Early recognition of these events, preferably before their
mistaking breathing for a seizure (or vice versa) could occurrence, is crucial for ensuring patient safety. 21,22 While
lead to inappropriate medical interventions or a failure to
respond promptly to an actual seizure. one-to-one observation is an effective way to monitor high-
risk patients, it is not feasible for all admitted patients. 23-26
Overall, despite the misclassifications, the numbers In addition, while video monitoring appears to be a viable
suggest a higher number of correct predictions across alternative solution, it can raise privacy concerns and
all classes than incorrect ones, indicating the promising incur substantial startup costs and resource investment
potential of the classifier. However, it is essential to identify to install cameras in multiple rooms. The utilization of
27
the clinical consequences of each type of misclassification Euler angle measurements through the Movella Xsens
to prioritize improvements in the classifier’s performance. sensors, in conjunction with machine learning algorithms,
offers a potential solution to these problems by enabling
4. Discussion the detection and classification of specific movements
This study, conducted with data collected using Xsens DOT in real time. While there are limitations to using these
sensors and analyzed utilizing PyCaret, showcases the sensors in practical settings that must still be addressed,
potential benefits of using artificial intelligence methods this study presents promising results and lays a foundation
in practical, real-life applications. The results indicate an for further research in this area. By accurately predicting
accuracy rate of over 89% in predicting the movements of ongoing motion, this novel approach can be incorporated
inpatients based on the provided data. The failure to achieve into a trigger alert system for the medical staff, allowing
a perfect accuracy rate from the LIGHTLGBM classifier for swift intervention before a fall or seizure occurs. This
is attributed to its difficulty in distinguishing between proactive approach will significantly reduce the risk of
breathing and seizures, with prediction values for these adverse events, lower complication rates, and ultimately
movements approximately 63% and 67%, respectively. This improve overall patient outcomes. 28-32
challenge likely stems from the absence of chest movement This study presents several limitations. The use of
during seizures, as opposed to the common occurrence a controlled environment with a limited number of
of head movements. With that being said, one potential mannequins to replicate patient movements does not
20
solution to this problem is to place an additional sensor fully capture the variability observed in actual patient
on the mannequin’s head to detect seizure movements. The populations. To better simulate real-world conditions,
data gathered from this sensor on seizures can be used as expanding the study’s scope to include a larger and
the evaluation metrics, while the sensor placed on the chest more diverse group of patients across various health-
will solely gather data on the other movements of interest. care settings would provide a more comprehensive
While promising, the results of this study do highlight understanding of sensor capabilities and algorithm
several additional limitations, beyond the one previously performance. In addition, sensors may have limitations in
mentioned, all of which fortunately are addressable accurately detecting subtle or intricate patient movements,

Volume 1 Issue 2 (2024) 139 doi: 10.36922/aih.2790