Page 67 - AIH-1-1

Artificial Intelligence in Health Advancing fetal health classification

for training purposes. This allocation ensures that a this study because it had been carefully curated and pre-
substantial portion of the data, representative of diverse processed, allowing us to evaluate the performance of this
scenarios, is utilized for training, contributing to the ML model on real-world fetal health data.
robustness of the model’s learning. The remaining 20% of
the samples in the test set provides a reliable benchmark 4.2. Data preprocessing
for evaluating the model’s performance on new and Before the experiments were conducted, data pre-
unseen instances, testing the model’s ability to generalize processing steps were performed to ensure the quality
data beyond the scope of the training data. This balanced and integrity of the dataset. These steps encompass
approach to dataset splitting enhances the reliability of the handling missing values, removing irrelevant features,
model’s assessments and fosters a more comprehensive and normalizing the data. Exploratory data analysis was
understanding of its predictive capabilities. also conducted to gain insights into the distribution and
By implementing this proposed methodology, this characteristics of the dataset, which guided our feature
research aimed to develop a robust and efficient ML model selection process.
for fetal health classification. The selected features, the 4.3. Feature selection
LightGBM classifier, and the training process were all
designed to achieve accurate and reliable predictions of fetal From the Ayres-de-Campos dataset, we selected 22
health. The following sections present the experimental features that were considered relevant to fetal health
results, discuss the model’s performance, and highlight assessment. These features included important parameters
the implications of the findings for improving fetal health such as fetal heart rate, uterine contractions, and maternal
assessment in clinical settings. blood pressure. These features were chosen based on their
known associations with fetal health and their availability
4. Experimental setup in the dataset. The selected features were represented as
numerical values and used as input to this ML model.
4.1. Dataset selection
The Ayres-de-Campos dataset titled “Sisporto 2.0: 4.4. Model training and evaluation
A program for automated analysis of cardiotocograms” The performance of several classification models on a
was utilized for this research. This dataset provided us with dataset was evaluated using different hyperparameters.
a diverse collection of CTGs containing information on The performance of the classification models is presented
various fetal health indicators. The dataset was suitable for in Figure 1. The LightGBM model achieved the best

Figure 1. Performances of the machine learning model in classification.

Figure 2. Numerical results of LightGBM model in fetal health classification.

Volume 1 Issue 1 (2024) 61 https://doi.org/10.36922/aih.2121