Page 140 - AIH-1-2

Artificial Intelligence in Health Movement detection with sensors and AI

falls and seizures as follows: Adult-sized and infant-sized (iv) Evaluation metrics: PyCaret-generated metrics such
mannequins were employed to represent a range of patient as accuracy, area under the curve (AUC), recall,
demographics, ensuring that the collected movement data precision, F1 score, and others are introduced. The
spanned various relevant physiologies for our algorithm’s definitions and significance of these metrics for model
predictive capabilities. The use of such mannequins allows evaluation are articulated, including the intricacies of
for consistent and repeatable movement simulations, how performance metrics, such as accuracy, precision,
which are crucial for machine learning applications. The recall, and F1 score, are calculated. Other important
rationale behind selecting the PyCaret machine learning metrics, such as the receiver operator characteristic
library is twofold. First, PyCaret’s low-code environment curve (ROC), AUC, Cohen’s Kappa, Matthews
significantly streamlines the development process, thereby correlation coefficient (MCC), and training time
facilitating a more efficient exploration of different (TT), are discussed, explaining each metric’s value
predictive models. Second, it offers a comprehensive suite range and its implications for the model’s predictive
of evaluation metrics and algorithms suitable for both performance.
binary and multiclass classification problems, making it
particularly well-suited for the complex task of classifying 2.1. Data collection
the nuanced movements indicative of potential falls or The data were collected using the Xsens DOT sensors
seizures. This adaptability and ease of use render PyCaret capable of capturing Euler angles in the X-, Y-, and Z-axes,
highly suitable for health-care settings, where rapid and also known as roll, pitch, and yaw, respectively, to depict
accurate decision-making is paramount for patient safety. real-time 3D orientation in space. In addition, these
sensors have demonstrated efficacy in capturing distinctive
This section also provides detailed descriptions of the 15
approaches used for data collection, preprocessing, and the data related to various patient movements. As illustrated
in Figure 1, one sensor was placed on the mannequin’s
setup of the machine learning model as follows: chest, specifically at the center of the sternum, to evaluate
(i) Data collection: This segment describes the use of the aforementioned six movements of interest. The sensor
Xsens DOT sensors to gather real-time motion data was securely affixed to the mannequin using duct tape
reflecting 3D orientations in space, which is capable arranged in a cross-shaped configuration. Subsequently, it
of detecting Euler angles in the X-, Y-, and Z-axes. was wirelessly connected through Bluetooth through the
It discusses the placement of the sensor and the Movella Dot App. The application allows for continuous
mechanics of its securement to the mannequin’s chest. streaming and data collection once initiated by the user.
The methodology elucidates the specific movements The sensors were activated and deactivated for each overall
imitated by the mannequins (e.g., breathing, seizures, movement of interest, with the collected data immediately
rolls, and falls) to collect diverse movement data while transferred to the device connected to the sensor. To
distinguishing between the use of adult and infant collect the data on breathing and seizures, an adult
mannequins for different movements. mannequin was used. Conversely, an infant mannequin
(ii) Data preprocessing: This section outlines the process was used to collect data on the remaining four movements.
of managing raw datasets, which involves the Data pertaining to breathing involved the mannequin
segregation of collected data into subsets correlating performing one full cycle of tidal volume inhalations and
to specific movements of interest. Focus is given to exhalations continuously for 3 min. Seizure data were
the significance of the Euler angle points in the X-axis collected by inducing a seizure in the mannequin for
and the quantification method for capturing distinct 10 min. For the collection of data on rolling to the side,
movement data, including simulated rolling and
falling off a bed by a mannequin.
(iii) Machine learning model setup: Details are provided
on the utilization of PyCaret, a supervised machine-
learning module for the study, emphasizing its
streamlined workflow and five key steps: setup,
compare models, analyze model, save model, and
prediction. The process of setting up PyCaret,
providing data, labeling the target, and ensuring
reproducibility through session IDs is described.
Detailed information about the dataset, including data
shape before and after transformations and division Figure 1. Depiction of the sensor placement and the respective angular
into training and test sets, is provided. motions. Image created using Inkscape

Volume 1 Issue 2 (2024) 134 doi: 10.36922/aih.2790