Page 66 - AIH-1-1

Artificial Intelligence in Health Advancing fetal health classification

2. Accelerations refer to the rapid increases in fetal heart and variability measures provide insights into the
rate, which are often associated with positive fetal health. physiological state, and the histogram features capture the
3. Fetal movement quantifies the overall movements of distribution patterns, offering a rich set of information for
the fetus, providing insights into its activity level. robust fetal classification algorithms.
4. Uterine contractions measure the intensity and
frequency of contractions, which are critical for 3.3. Model architecture
assessing stress on the fetus. To build the fetal health classification model, the LightGBM
5. Light decelerations refer to the temporary decreases in classifier was employed. LightGBM is a powerful and
the fetal heart rate, which indicate mild stress. efficient tree-based model that has demonstrated
6. Severe decelerations refer to the pronounced decreases exceptional performance in various classification tasks.
in the fetal heart rate, which signal more significant It provides fast training and prediction capabilities,
stress. which makes it an excellent choice for this research. The
7. Prolonged decelerations refer to the extended periods scikit-learn library’s LGBMClassifier implementation,
of reduced fetal heart rate, indicative of prolonged in combination with the default hyperparameters, was
stress. utilized.
8. Abnormal short-term variability refers to the irregular
variations in the fetal heart rate over short intervals, 3.4. Training process
which are associated with potential issues. During the training process, a 20-fold cross-validation
9. Mean value of short-term variability refers to the procedure was implemented to ensure reliable model
average magnitude of short-term variability, offering evaluation. This procedure involved dividing the dataset
a summary measure. into 20 subsets, performing training and evaluation
10. Percentage of time with abnormal long-term variability iterations, and aggregating the results. In addition, to
is the proportion of time exhibiting irregular long- address any class imbalance issues and enhance the
term variability, highlighting potential concerns. model’s performance, the synthetic minority over-
11. Histogram minimum refers to the minimum value in sampling technique (SMOTE) was applied to balance the
the distribution of fetal heart rate histogram. distributions of the different classes.
12. Histogram maximum refers to the maximum value in
the distribution of fetal heart rate histogram. The LGBMClassifier with the following settings was
13. Histogram number of peaks is the count of peaks in the employed:
histogram, providing insights into heart rate patterns. • Boosting type: Gradient Boosting Decision Tree
14. Histogram number of zeroes is the count of zero (gbdt)
values in the histogram, reflecting specific heart rate • Learning rate: 0.1
occurrences. • Maximum depth of trees: Unlimited (−1)
15. Histogram mode is the most frequently occurring • Minimum number of samples required in each leaf: 20
value in the histogram, indicative of a dominant heart • Minimum weight fraction of the sum total of weights:
rate. 0.001
16. Histogram mean is the average value in the histogram, • Minimum loss reduction required to make further
representing the central tendency of heart rate partition: 0.0
distribution. • Number of boosting iterations: 100
17. Histogram median is the middle value in the histogram, • Number of parallel threads for LightGBM: -1 (utilizing
offering an alternative measure of central tendency. all available threads)
18. Histogram variance refers to a measure of the spread or • Number of leaves in each tree: 31
dispersion of fetal heart rate values in the histogram. • Random state for reproducibility: 123
19. Histogram tendency describes the trend or shape of the • No regularization parameters (reg_alpha and reg_
fetal heart rate histogram. lambda) were applied
20. Fetal health is the target variable representing the • Silent mode enabled, only warnings will be displayed.
overall health classification of the fetus. The dataset was split into a training set, constituting
Each of these features contributes unique information 80% of the data, and a test set, containing the remaining
for a comprehensive fetal health assessment. For instance, 20%. The training set allowed for the evaluation of
accelerations and fetal movements are generally positive performance of the trained model on unseen data and
indicators, while decelerations, especially severe and the assessment of its generalization capability. Out of the
prolonged ones, may suggest stress. Uterine contractions total 2126 samples, approximately 1700 were designated

Volume 1 Issue 1 (2024) 60 https://doi.org/10.36922/aih.2121