Page 141 - AIH-1-2

Artificial Intelligence in Health Movement detection with sensors and AI

the infant mannequin was initially positioned in the the training environment and creates a transformation
supine position and then rolled approximately 90 degrees pipeline. It requires two mandatory parameters, “data”
to the left side and back to the original supine position, and “target,” and offers several optional parameters for
a procedure repeated 100 times. Similarly, the same customization. The user provides the data in a cohesive
procedure was performed for rolling to the right side. fashion with the target labeled appropriately, typically in
Data regarding falling off the bed from the left side were comma-separated values file (CSV) format. Since this is
collected by initially positioning the infant mannequin a classification model, the target is a categorical variable
in the supine position and subsequently rolling it beyond represented numerically (i.e., “Roll right” is 0, “Roll left” is
its left side until it fell off the bed, a procedure repeated 1, “Drop right” is 2, “Drop left” is 3, “Breathing” is 4, and
100 times. Given the near-identical nature of dropping “Seizure” is 5). The code base for the Google Colaboratory
the mannequin from its right side, this movement was not notebook is available at: Sensor_Classification.ipynb.
performed. Once the setup is executed successfully, it displays an

2.2. Data preprocessing information grid with experiment-level details (Table 1).
The session ID is a pseudo-random number (123 in this
The raw dataset for each respective movement, collected case) used as a seed for reproducibility in all functions
by the sensors, was immediately gathered following the throughout the PyCaret pipeline, ensuring consistent
completion of repeated movements. These datasets were results when running the same code with the same session
divided into six specific movements of interest, each ID. The target refers to the column in the dataset (the
ranging from approximately 18,000 to 35,000 data points CSV file) that will be predicted. In this case, the target is
containing the Euler angle points in the X-, Y-, and Z-axes. designated “Predict.” The target type specifies the nature
Of particular interest were the Euler angle points along the of the target variable, which in this case is “Multiclass,”
X-axes exclusively. To isolate the data for each movement, indicating that the target variable has multiple distinct
the raw dataset needed to be subdivided to capture data classes for multiclass classification. The original data
for every 100 movements. Approximately 200 data points shape shows the dimensions of the dataset before any
were needed to represent one complete movement. For transformations, with 579 rows and 203 columns.
example, the information of one complete roll to the left Similarly, the transformed data shape also has 579 rows and
side consists of 200 points, capturing the transition from 203 columns, indicating that the dataset was not modified
the starting supine position to rolling the mannequin to during the setup process. The transformed training set
its left side and back to the starting supine position. This shape indicates that the training dataset contains 405 rows
collection of 200 points was repeated to ensure a clear and 203 columns after preprocessing, which was used to
delineation of values for each of the 100 movements. This train the machine learning models. The transformed test
data segmentation process was similarly applied to the set shape indicates that the test dataset contains 174 rows
other movements of interest. To mimic the mannequin and 203 columns after preprocessing, which was used to
dropping from the right side, the values obtained from evaluate the performance of the trained models. Therefore,
dropping the mannequin from the left side were negated. a split of 70% for training and 30% for testing was used.
Once the data for each completed movement were
collected, it was transferred to one single Excel sheet for Table 1. Experiment setup details
further analysis. From the raw datasets, 100 movements No. Description Value
were collected for each roll to the right side, roll to the left 0 Session ID 123
side, and seizures. Ninety-five movements were collected 1 Target Predict
for each dropping off the bed from the left and right sides,
while 89 movements were collected for breathing. 2 Target type Multiclass
3 Original data shape (579, 203)
2.3. Pycaret setup 4 Transformed data shape (579, 203)
The classification module in PyCaret is a supervised 5 Transformed train set shape (405, 203)
machine learning module designed for classifying elements 6 Transformed test set shape (174, 203)
and aiming to predict categorical class labels that are discrete 7 Numeric features 202
and unordered. It can handle both binary and multiclass 8 Preprocess True
problems, finding applications in various scenarios. The 9 Imputation type Simple
typical workflow in PyCaret for classification consists of
five steps: setup, compare models, analyze model, save 10 Numeric imputation Mean
model, and prediction. The first step, “Setup,” initializes 11 Categorical imputation Mode

Volume 1 Issue 2 (2024) 135 doi: 10.36922/aih.2790