Optimizing Fetal Heartbeat Detection with Feature Selection in Medical Data Using Machine Learning Algorithms

Figure 1 presents a flowchart illustrating the research process, in which machine learning algorithms are applied both with and without feature selection to evaluate and enhance FHR detection. This approach aims to improve model efficiency while maintaining accuracy, with feature selection serving to identify and optimize the most relevant variables for effective FHR categorization.

Figure 1. The optimization of fetal heart rate detection through feature selection and machine learning algorithms [24]

2.1 Cardiotocography dataset

Figure 1 begins with the CTG dataset from the University of California, Irvine (UCI) Machine Learning Repository. This dataset consists of 2,126 recordings and includes 21 variables, including FHR, variability, acceleration, and other factors relevant to assessing fetal health.

The three target classes in the CTG dataset—Normal, Suspect, and Pathological—represent different levels of fetal risk. There is a significant imbalance between these classes, with 1,655 samples in the Normal category compared to only 295 in the Pathological category. This imbalance directly affects classification performance, especially for minority classes, making it an essential methodological consideration. To address this issue, stratified splitting was applied during data splitting to ensure consistent class proportions between the training and testing sets. In addition, evaluation metrics such as macro F1-scores were prioritized to accurately capture performance across classes rather than being dominated by the majority class.

2.2 Data preprocessing

The data underwent a preparation stage before being used for model training. This stage included:

1. Stratified data division (80/20): To maintain the same class distribution (1,655/295/176) in both the training and testing subsets, the dataset was divided into two parts.
2. Scaling (StandardScaler): All features were scaled using StandardScaler to normalize variance among predictors, allowing algorithms such as SVM and Logistic Regression to work optimally.

2.3 Separation of analysis paths

Once the data were ready, the process was divided into two parallel paths:

(1) Baseline evaluation path

On this path, all 21 features were used without feature selection. Before optimization, the main goal was to achieve the baseline performance of the model. The procedure included:

1. Data model, consisting of all 21 characteristics, was trained using three different methodologies: Random Forest, SVM, and Logistic Regression.
2. Baseline performance assessment, consisting of metrics such as recall, accuracy, precision, and F1-score to evaluate model performance before feature reduction.

(2) Feature selection and workflow optimization

Feature selection was employed to ensure the most informative subset of predictors to improve classification efficiency and accuracy. The ANOVA F-test evaluated the statistical significance of individual features, whereas RFE determined traits with negligible contributions. After identifying the ten most significant characteristics, the Random Forest, SVM, and Logistic Regression models were reconfigured using the reduced feature set. Subsequently, the effectiveness of the models was reevaluated to determine the impact of feature selection on classification results.

2.4 Final analysis

The final part of the flowchart shows the comparative analysis stage:

1. Comparing how well the optimized model (using only 10 characteristics) performed compared to the baseline model (using 21 features).
2. Analyzing the impact of feature selection on model accuracy, computational efficiency, and generalization comprehensively.

Figure 1 illustrates the systematic process of developing and optimizing an FHR classification model. This study focuses not only on model training but also evaluates the impact of feature selection on final performance. The two-path approach (baseline vs. optimization) helps ensure that performance improvements were not coincidental, but rather the result of appropriate feature selection.

2.5 Feature selection method

Feature selection was employed to ensure the most informative subset of predictors to improve classification efficiency and accuracy. The ANOVA F-test evaluated the statistical significance of individual features, while RFE determined traits with negligible contributions. After identifying the ten most significant features, the Random Forest, SVM, and Logistic Regression models were retrained using the reduced feature set. The ANOVA F-test was chosen as an effective univariate filter method to measure statistical differences in feature distributions among fetal health categories. This method provides quick and objective rankings for numerical variables, making it suitable for an initial screening stage. RFE was chosen as a wrapper method to complement the screening approach by considering multivariate interactions that cannot be captured by the ANOVA F-test. Compared to other alternatives, such as Mutual Information or SHapley Additive exPlanations (SHAP), the ANOVA F-test, and RFE, were preferred due to (1) lower computational cost, (2) suitability for limited sample sizes such as CTG datasets, and (3) better alignment with the study objective to balance performance and readability. SHAP, while highly interpretable, requires substantial computational resources and is generally used for post-analysis explanation rather than feature selection. Mutual Information was not selected because it is less effective with correlated features, which are commonly found in CTG measurements.

2.5.1 Analysis of Variance F-test (filter method)

For statistical purposes, the relationship between each numerical characteristic and the categorical fetal health outcomes was analyzed using the ANOVA F-test. A reduced feature subset was initially formed by selecting features with the largest F-statistic values using SelectKBest from Scikit-learn.

2.5.2 Recursive Feature Elimination (wrapper method)

RFE was implemented using Logistic Regression as an estimator. This algorithm repeatedly removed features with the lowest importance until a predetermined number of selected features was reached. As a wrapper method, RFE evaluated feature combinations in conjunction with the algorithm, enabling the selection of a more representative multivariate feature subset.

2.5.3 Classification algorithm

This study utilized three machine learning algorithms, Random Forest, Logistic Regression, and SVM, to identify fetal health problems. Grid search with 5-fold stratified cross-validation was used to adjust the hyperparameters of SVM and Random Forest. With this optimization, we ensured that the selected parameters were robust, reduced the possibility of overfitting, and helped achieve the goal of improving model performance.

(1) Logistic Regression

Logistic Regression was used as a baseline classifier due to its ease of interpretation and efficiency. The model employed L2 regularization penalty with max_iter = 1000 to ensure convergence.

(2) SVM

Instead of using default parameters, SVM with the RBF kernel was optimized using grid search.

Explored grid parameters:

• C: {0.1, 1, 10, 100}
• gamma: {1e-3, 1e-2, 1e-1, ‘scale’}
• kernel: {‘rbf’}

Optimal parameters:

• C = 10
• gamma = 0.01
• kernel = 'rbf'

These adjusted parameters improved class separation and enhanced performance for minority fetal-health classes.

(3) Random Forest

Random Forest hyperparameters were adjusted to increase tree diversity and reduce overfitting.

Explored grid parameter:

• n_estimators: {100, 200, 300, 500}
• max_depth: {None, 10, 20, 30}
• min_samples_split: {2, 5, 10}
• min_samples_leaf: {1, 2, 4}
• max_features: {‘sqrt’, ‘log2’}

Optimal parameters:

• n_estimators = 300
• max_depth = 20
• min_samples_split = 5
• min_samples_leaf = 2
• max_features = 'sqrt'

This optimization produced a more stable ensemble with better generalization.

2.5.4 Evaluation metrics

The metrics included a confusion matrix, accuracy, precision, recall, macro/micro F1-score, and algorithm performance evaluation. Due to the class imbalance, greater emphasis was placed on the macro F1-score and per-class recall, especially for the minority classes (Suspect and Pathological), which are clinically critical.

2.5.5 Feature selection implementation

This software identified the ten most significant features from the normalized data with the ANOVA F-test and RFE. The selection approach sought to improve generalization, minimize redundancy, and augment computing efficiency. Eleven attributes were deliberately chosen. An initial assessment was conducted by analyzing several k values, including 5, 8, 10, 12, and 15 characteristics, to determine the ideal number of features. The ANOVA F-test and RFE were utilized for this purpose. The results showed that k = 10 consistently produced the best balance between computational cost, accuracy, and macro F1-score (especially for minority classes). Retaining fewer than 10 features resulted in the loss of important clinical information, while selecting more than 10 features increased redundancy and computation time without meaningful performance improvement. Additionally, the decision aligns with prior CTG feature-selection studies, in which 8–12 features are commonly retained to achieve interpretability, avoid overfitting, and maintain relevant clinical variables. Therefore, the selection of 10 features is methodologically justified and based on in empirical testing and consistency with the existing literature.

2.5.6 Model training without feature selection

To evaluate the initial performance of each classifier using all 21 characteristics, model training without feature selection was performed as a baseline. Unlike the previous version, this revised baseline integrated optimized SVM and Random Forest hyperparameters to ensure a fair comparison between models before and after feature selection (Table 1).

Table 1. Model training without feature selection

In this first phase, all classifiers were trained on all scaled features using optimized hyperparameters for SVM and Random Forest. The test set was used to generate predictions, and performance was evaluated by computing the F1-score, recall, accuracy, and precision. The optimized configuration ensured that the baseline reflected the performance capability of each algorithm without relying on default parameters.

2.5.7 Model training with feature selection

The top ten features chosen using an ANOVA F-test and RFE were used to train the model with feature selection. To ensure consistency and determine how feature reduction affected model performance, the dataset was reprocessed using improved SVM and Random Forest configurations (Table 2).

Table 2. Model training with feature selection

All models were trained using the identically optimized hyperparameters used during the baseline phase, incorporating a condensed 10-feature dataset. This ensured that performance variances were only attributable to feature selection rather than alterations in parameters. Performance metrics were evaluated to determine whether feature reduction improved accuracy, computational efficiency, and class-balance sensitivity.

2.6 Feature optimization in machine learning

Feature optimization is a critical pre-modeling step in machine learning that improves algorithm performance through the selection of the most relevant and informative features. By eliminating redundant or low-contributing variables, this process reduces model complexity, mitigates overfitting, and improves computational efficiency. Common feature optimization approaches include filter, wrapper, and embedded methods. When applied correctly, these techniques enable machine learning models to produce more accurate and interpretable predictive results.

2.6.1 Importing algorithm library

Importing several important libraries to support the entire process, this application was the initial stage in data processing and analysis using machine learning techniques. Data manipulation was accomplished using libraries such as pandas and numpy, while data visualization was executed using matplotlib.pyplot, and seaborn. Features were standardized using the sklearn.model_selection and sklearn.preprocessing modules, and the data were partitioned into training and test sets. Two common techniques for feature selection are RFE and SelectKBest, which use the f_classif statistical approach. Various leading methods, including XGBoost, Random Forest, Logistic Regression, and SVM, were employed in the classification phase. Finally, to assess model performance, the sklearn.metrics module employed evaluation measures such as accuracy, confusion matrix, and classification report (Table 3).

Table 3. Libraries, modules, and algorithms used in the optimization workflow

2.6.2 Data loading and preparation

Data loading and preparation is the initial stage in machine learning that aims to load and prepare data so that it can be used by the model. This step involved retrieving data from sources, removing errors or gaps, and transforming the data through normalization and dividing it into training and testing datasets. Data quality is extremely important since it substantially affects the accuracy of model predictions. Various algorithms are available (Table 4).

Table 4. Data loading and preparation workflow

This program aimed to load and prepare data from the fetal_health.csv dataset for fetal health condition classification analysis. After ensuring the file was available, the program displayed the data structure and distribution to understand the initial condition of the dataset. The data were then divided into features (X) and targets (y), and further divided into training and test datasets using stratification to maintain a balanced class proportion. The features in the dataset were normalized using StandardScaler, which is crucial for improving the performance of models such as SVM and Logistic Regression.

2.6.3 Baseline data

Initial data collected before treatment or intervention function as a reference point for comparing changes or effects of a process, program, or model. In this study, baseline data are important for determining the initial condition of the research object so that the results of the analysis can be measured objectively.

Table 5 presents a complete list of the dataset's numerical attributes, along with the target variable fetal_health, all of which have a uniform data type and complete values with no nulls. These features include FHR characteristics, variability patterns, and other histogram-based markers, which are clinical and statistical metrics derived from CTG signals. The completeness and consistency of the data indicate that the dataset is in optimal condition for machine learning analysis and modeling without requiring additional data-cleaning procedures.

Table 5. Baseline data algorithm

The dataset comprised 2,126 observations and 21 numerical variables, all complete with no missing values. These features capture FHR characteristics and histogram-based descriptors generated from CTG recordings. Table 6 shows that the distribution of fetal health is uneven. Most instances are classified as Normal (1.0), while there are far fewer samples in the Suspect (2.0) and Pathological (3.0) classifications. This pronounced class imbalance has methodological implications, requiring careful interpretation of classification performance, particularly for minority classes, and underscores the importance of model evaluation strategies that account for unequal class representation.

Table 6. Target class distribution

2.6.4 Model training without feature selection as baseline

Feature selection-free model training was performed to establish a baseline as a reference for objectively evaluating the impact of the optimization process. At this stage, all available features were included in the training phase, allowing each algorithm to learn from the complete feature set without any prior dimensionality reduction. The performance of this baseline serves as a reference point to determine whether subsequent feature selection enhances model accuracy, computational efficiency, or generalization ability in classifying fetal health conditions (Table 7).

Table 7. Model training without feature selection as baseline

The basic procedure consists of three stages: (1) Training the algorithm with the scaled training dataset (X_train_scaled, y_train); (2) Utilizing the scaled test dataset (X_test_scaled) for predictions; and (3) Evaluating performance through various metrics, including accuracy and a comprehensive classification report (precision, recall, and F1-score for each class). Analysis of Random Forest, SVMs, and Logistic Regression was performed under the same conditions to ensure objectivity. These initial results form the basis for evaluating feature selection in the next optimization phase.

2.6.5 Feature selection using Analysis of Variance F-test and Recursive Feature Elimination

During feature selection implementation, the ten most relevant features from the 21 characteristics in the CTG data were identified using the ANOVA F-test and RFE methods. The objective was to simplify the model, lower the risk of overfitting, and improve accuracy and computational efficiency in FHR classification (Table 8).

Table 8. Features of the Random Forest model algorithm

This software employed the ANOVA F-test and RFE to choose the top 10 features from the scaled data. RFE uses the Logistic Regression model to systematically exclude less significant characteristics, whereas the ANOVA F-test selects features based on the robustness of their statistical association with the objective. The selected features aimed at improving the efficiency and accuracy of the classification model, as detailed in the findings.

2.6.6 Model training with feature selection

Feature selection training using RFE results was performed by training classification models such as Logistic Regression, SVM, and Random Forest only on the top 10 features. Classification of FHR problems should be carried out with consideration for improving accuracy while reducing model complexity and the risk of overfitting (Table 9).

Table 9. Model training with feature selection algorithms

Significance of the 0.5% accuracy gain: The observed improvement in Random Forest classification from 92.72% to 93.19% is relatively small, and because the experiment was conducted using a single stratified train-test split, this improvement cannot be interpreted as statistically significant. To make such a claim, repeated k-fold cross-validation combined with a statistical test such as the McNemar test or a paired t-test on repeated runs is necessary.

This study demonstrates that feature selection can enhance model efficiency and interpretability in CTG-based fetal health classification, although its impact varies across algorithms. Random Forest shows the most substantial improvement, with accuracy improving from 92.72% to 93.19% and the macro F1-score rising from 86% to 87%, indicating that reducing redundant features enhanced its predictive capability. In contrast, Logistic Regression and SVM exhibited a slight decrease in performance after feature reduction, suggesting that feature selection is model-dependent rather than universally advantageous. Overall, feature selection remains a useful approach for simplifying models and improving clinical applicability, particularly when combined with robust algorithms such as Random Forest.