An improved ensemble model of hyper parameter tuned ML algorithms for fetal health prediction

Abstract

Fetal health is a critical concern during pregnancy as it can impact the well-being of both the mother and the baby. Regular monitoring and timely interventions are necessary to ensure the best possible outcomes. While there are various methods to monitor fetal health in the mother's womb, the use of artificial intelligence can improve the accuracy, efficiency, and speed of diagnosis. In this study, we propose a robust ensemble model called ensemble of tuned Support Vector Machine and ExtraTrees (ETSE) for predicting fetal health. Initially, we employed various data preprocessing techniques such as outlier rejection, missing value imputation, data standardization, and data sampling. Then, seven machine learning classifiers including Support Vector Machine, XGBoost, Light Gradient Boosting Machine, Decision Tree, Random Forest, ExtraTrees, and K-Neighbors were implemented. These models were evaluated and then optimized by hyperparameter tuning using the grid search technique. Finally, we analyzed the performance of our proposed ETSE model. The performance analysis of each model revealed that our proposed ETSE model outperformed the other models with 100% precision, 100% recall, 100% F1-score, and 99.66% accuracy. This indicates that the ETSE model can effectively predict fetal health, which can aid in timely interventions and improve outcomes for both the mother and the baby.

Introduction

Pregnancy is a joyous and memorable time for women, but it is also a period that requires special care for both the mother and fetus. Abnormal fetal growth poses significant risks to maternal health and mortality, with approximately 810 women dying every day during pregnancy according to WHO [1]. The maternal mortality ratio is higher in underdeveloped countries, highlighting the crucial need for fetal health monitoring. Doctors recommend regular tests to monitor the condition of the fetus, with Cardiotocography (CTG) being a widely used technique for assessing fetal and maternal health [2]. CTG continuously records the fetal heart rate (FHR) and uterine contractions (UC) via an ultrasound transducer located on the mother's abdomen [3]. The CTG data includes 21 attributes that obstetricians use to determine fetal health status, enabling early detection and treatment of fetal distress before it becomes severe [4, 5]. The emergence of artificial intelligence has facilitated further advancements in fetal health monitoring [6, 7, 8, 9, 10–11]. However, accurately predicting fetal health conditions remains a major challenge in the field of machine learning.

In our work, we utilized the UCI CTG dataset to improve the accuracy of fetal health prediction. We proposed a highly accurate ensemble of tuned SVM and ExtraTrees (ETSE) model that outperforms existing literature. We applied various data preprocessing techniques such as outlier rejection, missing value imputation, data standardization, and data sampling to enhance the quality of the data fed into the models. Furthermore, we implemented hyper-parameter tuning using the Grid Search Technique to improve the accuracy of each conventional classifier used in fetal health prediction. Finally, we ensembled different models to create a benchmark model for accurate fetal health prediction. Our proposed approach demonstrates significant improvements in predicting fetal health conditions, indicating the potential of artificial intelligence in improving maternal and fetal health outcomes.

The paper is structured as follows: Sect. 1 provides an overview of the study, including the purpose and significance of the research. Section 2 reviews related works in the field and identifies gaps addressed by our study. Section 3 describes the materials and methods used in our research, including data sources, study design, and statistical analysis. Section 4 presents the results of the study in a clear and concise manner. Finally, Sect. 5 offers a comprehensive discussion of the results, including a critical evaluation of the study, its implications for the field, and recommendations for future research.

Related work

Artificial intelligence (AI) is currently one of the most popular topics in healthcare automation, with extensive use in disease detection, diagnosis, and medical treatment. A review of the literature shows that numerous studies have been conducted on the prediction of various diseases, such as diabetes, heart disease, fetal health disorders, and many more. In a study by Amzad Hossen et al. [12], the performance of three supervised machine learning models, namely RF, (DT, and logistic regression (LR), was investigated in heart disease analysis and prediction. The results showed that LR achieved the highest accuracy score of 92.10%. It is noted that this study did not explore other forms of conventionally supervised machine learning. However, Abdul Saboor et al. [13] used nine classifiers of machine learning including AdaBoost (AB), LR ET, MNB (Multinomial Naive Bayes), CART (Classification and Regression Trees), SVM, LDA (Linear Discriminant Analysis), RF, and XGB to predict human heart diseases. The experimental results revealed that the accuracy of the prediction classifiers improved with hyperparameter tuning and achieved notable results with data standardization and machine learning hyperparameter tuning. One such researcher is Sundar [14], who calculated precision, recall, and F-score for the commonly used unsupervised clustering method k-means clustering in the prediction of fetal health diseases. However, the performance of this method was not satisfactory. To improve upon these results, Sundar et al. [3] implemented a model-based CTG data classification system using a supervised artificial neural network (ANN), which yielded significantly improved performance. The values of precision, recall, and F1 score all showed notable improvements. The authors Ocak and Ertunc [15] used adaptive neuro-fuzzy inference techniques to classify CTGs (ANFIS), while Ocak developed a classification algorithm based on SVM and genetic algorithms (GA) [16]. Muhammad Arif et al. [4] proposed a RF classifier to distinguish between normal, suspicious, and pathological patterns. The results showed that the random forest classifier successfully identified these patterns, achieving an overall classification accuracy of 93.6 percent. Similarly, Tomáš Peterek et al. [17] employed the RF method and achieved exceptional performance, achieving 94.69% accuracy in classifying the data. It is important to note that the author only utilized one supervised machine learning method and did not explore data processing, hyperparameter tuning, or other machine learning models. On the other hand, Abolfazl Mehbodniya et al. [18] applied four supervised machine learning models, namely SVM, RF, multi-layer perceptron, and KNN, to predict the health state of the fetus. The RF algorithm achieved an accuracy of 94.5% on the CTG dataset. However, the accuracy is still considered low. To improve upon this, Nabillah Rahmayanti et al. [19] emphasized the importance of outlier removal, multicollinearity removal using VIF, data balancing, data scaling, and standardization. The authors applied seven algorithms, namely Artificial Neural Network, Long-Short Term Memory (LSTM), XGB, SVM, KNN, LGBM, and RF, to predict fetal health and compared their performance. The LGBM algorithm performed well across all seven scenarios. Recently, Md Takbir Alam et al. [20] conducted an analysis of multiple machine learning models, including RF, LR, DT, SVM, voting classifier, and K-nearest neighbor, to classify fetal health using the CTG dataset. The results of this study showed that the RF classifier achieved the highest accuracy of 97.51%, surpassing the accuracy achieved by previous studies. A summary of the performance comparison of different fetal health classification study is provided in Table 1.

Table 1. Summary of findings from the litteraure review

References	Models	Dataset	Hyper parameters tuning	Ensemble learning	Metrics (%)
[15]	ANFIS	Own (normal, suspect, pathological)	No	No	Accuracy = 91.6
[16]	SVM and GA	Own (normal and abnormal)	No	No	Accuracy = 98
[4]	RF	Own (normal, suspect, pathological)	No	No	Accuracy = 93.6
[17]	RF	Own (normal, suspect, pathological)	No	No	Accuracy = 94.69
[18]	RF	Own (normal, suspect, pathological)	No	No	Accuracy = 94.5
[20]	RF	UCI Public Dataset (normal, suspect, pathological)	No	No	Accuracy = 97.51

Researchers have employed advanced machine learning and data preprocessing techniques for fetal health detection. While most used their datasets, reference [20] achieved 97.51% accuracy using the publicly available UCI dataset. Yet, hyperparameter techniques and ensemble learning remain unexplored.

Our study introduces the ETSE model, leveraging hyperparameter tuning in various supervised machine learning models and analyzing their ensemble effect. We employ ExtraTrees classifiers, a novel approach for this dataset, and find that the ETSE (tuned ET and SVM ensemble) performs best. Our contributions include high accuracy in fetal health classification, data preprocessing enhancements (outlier rejection, missing value imputation, data standardization, and data sampling), Grid Search Techniques for Hyperparameters tuning, introducing ExtraTrees, and showcasing the power of ensemble learning to improve our model. This work adds to the literature on machine learning-based fetal health classification.

Materials and methodology

Our work provides an efficient prediction of fetal health condition while in utero, which is critical for the well-being of both the mother and the unborn child. The methods used in this study are thoroughly described in the following sections.

Dataset

Our experiment utilized the publicly available cardiotocography (CTG) dataset from the UCI Machine Learning repository [21]. This dataset contains 2126 records of features extracted from cardiotocograph exams, which were classified into three categories based on the assessment of three expert obstetricians using 23 different attributes. The attributes used in the measurement of fetal heart rate (FHR) and uterine contractions (UC) on CTG are listed in Table 2. The dataset is imbalanced, with 1655 records classified as normal, 295 as suspect, and 176 as pathological.

Table 2. CTG dataset attributes used in our model

Variable symbol	Variable description	Range
LB	FHR baseline (beats per minute)	(160–106)
AC	Accelerations per second	(0–0.019)
FM	Fetal movement (number of fetal movement)	(0–0.481)
UC	Uterine contractions (number of uterine contractions per second)	(0–0.015)
DL	Light decelerations (number of light decelerations per second)	(0–0.015)
DS	Severe decelerations (number of severe decelerations per second)	(0–0.001)
DP	Prolonged decelerations (number of prolonged decelerations per second)	(0–0.005)
ASTV	Abnormal short-term variability (percentage of time with abnormal short-term variability)	(12–87)
MSTV	Mean value of short-term variability	(7–0.2)
ALTV	Percentage of time with abnormal long-term variability	(0–91)
MLTV	Mean value of long-term variability	(0–50.7)
Width	Width of FHR histogram	(3–180)
Min	Minimum of FHR histogram	(50–159)
Max	Maximum of FHR histogram	(122–238)
Nmax	Histogram peaks	(0–18)
Nzeros	Histogram zeros	(0–10)
Mode	Histogram mode	(60–187)
Mean	Histogram means	(73–182)
Median	Histogram median	(77–186)
Variance	Histogram variance	(0–269)
Tendency	Histogram tendency	1
CLASS	FHR pattern class code (1 to 10)	(1–10)
NSP	Fetal health (Fetal state class code, N = normal, S = Suspected, P = Pathological)	(1-Normal, 2-Needs Reassurance, 3-Pathological)

Data pre-processing

Data processing plays a crucial role in improving model performance, especially when dealing with real-time datasets containing missing values and noisy data. In our study, we carefully checked for missing data, finding none. To handle outliers, we utilized standard deviation to remove them effectively. However, we encountered the challenge of imbalanced data, which could lead the model to overlook the minority class. To address this issue, we employed the Random Over Sampler (ROS) technique [22], duplicating data from the minority class to balance the dataset. This approach helps mitigate overfitting on skewed classes and enhances overall model performance. Additionally, we applied standardization (S) to rescale the distributions, achieving zero mean and one standard deviation (as shown in Eq. (1)), which reduced data distribution skewness.

1

$$S\left(x\right)=\frac{x-\overline{x}}{\sigma }$$where, X refers to the n-dimensional instances of the feature vector. As a result of standardizing, it results in a more stable model that is less influenced by variables, fits faster, and performs more consistently. Therefore, scaling is a crucial aspect of data pre-processing [23]. Once the data was standardized, we proceeded to split the dataset into training and testing sets, using a ratio of 70% for training and 30% for testing.

Grid search and hyper-parameter tunning

Hyperparameter tuning is crucial for optimizing machine learning models and achieving better performance [24]. The most common method for hyperparameter tuning is grid search (GS). In GS, hyperparameter values are explored in discrete grids, and the combinations with the best model performance are selected. The 'GridSearchCV' function in the sklearn library facilitates this process. Grid search is widely preferred for its ease of implementation, reliability, and efficiency [25]. In our work, we applied grid search to fine-tune all fundamental models, ensuring our model achieved optimal performance.

Ensemble learning

Ensemble learning combines machine learning models to create a more powerful model, addressing their shortcomings and boosting traditional machine learning performance [26]. The voting classifier is one such ensemble technique that integrates various models and predicts based on majority voting [27]. Different combination rules, like majority voting or probabilistic product, merge the outputs of these base predictors [26]. Two voting techniques are hard voting and soft voting. Hard voting selects the most commonly predicted class among the base models, while soft voting averages the highest probabilities from individual predictors. In our work, we utilized hard voting, and Algorithm 1 demonstrates the process.

Proposed model

We have proposed a robust ETSE model that outperforms previous studies. The entire procedure is presented in Algorithm 1. In this study, we applied seven ML models, including SVM, XGB, LGBM, DT, RF, ET, and K-Neighbors, to the training data. We applied Grid Search to each model to search for optimal parameter values. The optimal parameters obtained by Grid Search are presented in Table 3. By tuning the hyperparameters, we selected the most suitable model, which was then tested with the test dataset. We evaluated the model using the confusion matrix and calculated precision, recall, F1-score, and accuracy. The entire process is illustrated in Fig. 1. In the final stage, we used a voting classifier to create an ensemble of two or three models based on their performance. We executed the ensemble of LGBM and SVM, XGB and SVM, Extra Trees and SVM, RF and SVM, DT and SVM, ET and LGBM, and analyzed the performance of each ensemble model. We found that the ensemble of Extra Tree and SVM was the most successful and proposed model. We also examined the ensemble of ET, SVM, and RF. The proposed ETSE model is illustrated in Fig. 2.

Table 3. Tuned parameters

Algorithm	Parameters
SVM	C = 1000, decision_function_shape = 'ovo', gamma = 0.9, probability = True
XGB	max_depth = 12, n_estimators = 200
LGBM	boosting_type = 'goss', max_depth = 200, random_state = 100, silent = True, metric = 'None', n_jobs = 4, num_leaves = 20, n_estimators = 100
DT	ccp_alpha = 0.001, criterion = 'entropy', max_depth = 12, max_features = 'log2'
RF	max_features = 'auto', n_estimators = 100, max_depth = 12, criterion = 'entropy'
ET	random_state = 1, n_estimators = 340, max_features = None, max_depth = 25, criterion = 'gini'
KNN	n_neighbors = 3, p = 1, weights = 'distance'

[See PDF for image]

Fig. 1

Proposed methodology

[See PDF for image]

Fig. 2

Proposed ensemble of tuned SVM and ET (ETSE) model

XXX

Model performance analysis

In evaluating our proposed model, we employ both quantitative and qualitative methods. Qualitative evaluation involves visually comparing and assessing image classification results [26]. For quantitative evaluation, we use a confusion matrix and performance measures such as accuracy, precision, recall, and F1 score. These measures are defined using the following equations-

2

$$\mathrm{Accuracy}=\frac{\sum TP+\sum TN}{\sum TP+\sum TN+\sum FP+\sum FN}\ast 100$$

3

$$\mathrm{Precision}=\frac{\sum TP}{\sum TP+\sum FP}\ast 100$$

4

$$\mathrm{Recall}=\frac{\sum TP}{\sum TP+\sum FN}$$

5

$$\mathrm{F}1\mathrm{score}=2\ast \left(\frac{Precision\ast Recall}{Precision+Recall}\right)\ast 100$$where: TP denotes the true positive; TN is true negative; FP denotes the false positive; FN refers to the false negative.

Result analysis

In our experiment, we successfully predicted fetal health using seven different machine learning models, namely SVM, XGB, KGBM, DT, RF, ExtraTrees, and K-Neighbor. Tuned parameters for each model were obtained through a grid search, and the confusion matrix for each model was presented in Fig. 3. The number of misclassifications for the tuned SVM, XGB, KGBM, DT, RF, ExtraTrees, and K-Neighbor models are 9, 22, 22, 41, 22, and 52, respectively. In Table 4, the overall precision, recall, F1 score, and accuracy of each model are calculated. The tuned SVM model yields the highest precision, recall, and F1 score of 99% and an accuracy of 99.39%, while K-Neighbor has the lowest precision, recall, F1 score, and accuracy of 97%, 97%, 96%, and 96.51%, respectively.

[See PDF for image]

Fig. 3

Confusion matrix of each model

Table 4. Overall performance analysis of models

Model	Precision	Recall	F1-score	Accuracy
SVM	99	99	99	99.39
XGB	99	99	99	98.52
LGBM	99	99	99	98.52
Decision tree	97	97	97	97.24
Random forest	98	98	98	98.45
Extra trees	99	99	99	98.79
KNeighbors	97	97	96	96.51
LGBM + SVM	100	100	100	99.53
XGB + SVM	100	100	100	99.53
RF + SVM	100	100	100	99.53
DT + SVM	99	99	99	98.66
ET + LGBM	99	99	99	99.06
DT + ET	99	99	99	98.52
XGB + ET	99	99	99	99.06
XGB + LGBM	99	99	99	98.86
SVM + ET + RF	99	99	99	98.86
ETSE model	100	100	100	99.66

To improve the classification accuracy, we created ensemble models by combining the top-performing classifiers. We tested ten ensemble models, including LGBM + SVM, XGB + SVM, ExtraTrees + SVM, RF + SVM, DT + SVM, ET + LGBM, DT + ET, XGB + ET, XGB + LGBM, and SVM + ET + RF. Figure 3 displays the confusion matrix for each ensemble model. The LGBM + SVM, XGB + SVM, RF + SVM, and RF + SVM ensemble models showed the most significant improvement in precision, recall, and F1 score, resulting in an accuracy of 99.53%, which is higher than SVM's accuracy of 99.39%. However, other ensemble models like DT + SVM, ET + LGBM, DT + ET, XGB + ET, and XGB + LGBM did not outperform the tuned SVM. The proposed Ensemble of Tuned SVM and ExtraTrees (ETSE) model showed a significant improvement in precision, recall, and F1 score, achieving 100% precision, recall, and F1 score, and accuracy of 99.66%. The ensemble of the best three classifiers (SVM + ET + RF) could not improve the classification accuracy of this dataset. In conclusion, our proposed ETSE model has been proved as the best model for fetal health prediction.

Discussion and conclusion

The performance of individual models (SVM, XGB, LGBM, DT, RF, ET, and K-Neighbors) has significantly improved compared to the previous study [20], as shown in Table 5. Two main reasons account for the increased accuracy of the models: first, data preprocessing operations such as outlier rejection, missing value imputation, data standardization, and data sampling; second, hyperparameter tuning of the models using GridSearchCV. The SVM classifier achieved the highest accuracy of 99.39%, with 99% precision, recall, and F1-score. Furthermore, ensemble of different models was investigated to further enhance accuracy. The ET and SVM ensemble achieved the highest accuracy of 99.66%, with 100% precision, recall, and F1-score. The application of ensemble learning and hyperparameter tuning, which were not used in the reference [20], played a critical role in achieving these results. This proposed ETSE model can be a valuable tool in detecting fetal health in the womb and reducing fetal deaths, resulting in a healthier infant.

Table 5. Comparisons of our work

Model	Precision (%)	Recall (%)	F1-score (%)	Accuracy (%)
SVM [20]	97	97	97	96.57
RF [20]	98	98	98	97.51
DT [20]	96	96	96	95.70
K-Neighbour [20]	91	90	90	90.20
SVM (our paper)	99	99	99	99.39
RF (our paper)	98	98	98	98.45
DT (our paper)	97	97	97	97.24
K-Neighbour (our paper)	97	97	96	96.51
Our proposed ETSE	100	100	100	99.66

This work has a few limitations that should be acknowledged. Firstly, the model has not been tested in real-time clinical settings, which may limit its generalizability and practical application. Secondly, due to the unavailability of different datasets, the model has not been validated on other datasets, which may affect its robustness and reliability. Another limitation is the scarcity of references in this field, which may affect the comprehensiveness of the study. Finally, the deployment of the model is not discussed in our work. There may be additional technical and logistical challenges that could arise in practical deployment. To address these limitations, future work could involve extensive validation of the model on patients at nearby hospitals and clinics. Additionally, introducing new machine learning concepts such as weighted ensembles or other advanced techniques may further enhance the model's accuracy and performance.

Author contributions

MSHT: Implementation and drafting; SA: Writing, editing and correcting.

Funding

This research received no external funding.

Data availability

The data will be provided upon request.

Declarations

Conflict of interest

The authors declare no conflict of interest.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. O'Neill, E; Thorp, J. Antepartum evaluation of the fetus and fetal well being. Clin Obstet Gynecol; 2012; 55, 3 pp. 722-730. [DOI: https://dx.doi.org/10.1097/GRF.0b013e318253b318.PMID:22828105;PMCID:PMC3684248] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/22828105][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3684248]

2. Ocak, H; Ertunc, H. Prediction of fetal state from the cardiotocogram recordings using adaptive neuro-fuzzy inference systems. Neural Comput Appl; 2012; 23, 6 pp. 1583-1589. [DOI: https://dx.doi.org/10.1007/s00521-012-1110-3]

3. Sundar, C. An analysis on the performance of K-means clustering algorithm for cardiotocogram data clustering. Int J Comput Sci Appl; 2012; 2, 5 pp. 11-20. [DOI: https://dx.doi.org/10.5121/ijcsa.2012.2502]

4. Arif, M. Classification of cardiotocograms using random forest classifier and selection of important features from cardiotocogram signal. Biomater Biomech Bioeng; 2015; 2, 3 pp. 173-183. [DOI: https://dx.doi.org/10.12989/bme.2015.2.3.173]

5. Ingemarsson, I. Fetal monitoring during labor. Neonatology; 2009; 95, 4 pp. 342-346. [DOI: https://dx.doi.org/10.1159/000209299] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19494555]

6. News, FIGO. Report of the FIGO study group on the assessment of new technology. Int J Gynecol Obstet; 1997; 59, 2 pp. 169-173. [DOI: https://dx.doi.org/10.1016/S0020-7292(97)00208-7]

7. Miao, GJ; Miao, KH; Miao, JH. Neural pattern recognition model for breast cancer diagnosis. Multidiscip J Sci Technol J Select Areas Bioinform; 2012; 2, 8 pp. 1-8.

8. Miao, KH; Miao, JH; Miao, GJ. Diagnosing coronary heart disease using ensemble machine learning. Int J Adv Comput Sci Appl; 2016; 7, 10 pp. 30-39.

9. Miao, KH; Miao, GJ. Mammographic diagnosis for breast cancer biopsy predictions using neural network classification model and receiver operating characteristic (ROC) curve evaluation. Multidiscip J Sci Technol J Select Areas Bioinform; 2013; 3, 9 pp. 1-10.

10. Miao, JH; Miao, KH; Miao, GJ. Breast cancer biopsy predictions based on mammographic diagnosis using support vector machine learning. Multidiscip J Sci Technol J Select Areas Bioinform; 2015; 5, 4 pp. 1-9.

11. Ayres-de-Campos, D; Bernardes, J; Garrido, A; Marques-de-Sá, J; Pereira-Leite, L. Sisporto 2.0: a program for automated analysis of cardiotocograms. J Matern Fetal Med; 2000; 9, 5 pp. 311-318.1:STN:280:DC%2BD3M7lt1KktA%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/11132590]

12. Hossen, MDA; Tazin, T; Khan, S; Alam, E; Sojib, HA; Monirujjaman Khan, M; Alsufyani, A. Supervised machine learning-based cardiovascular disease analysis and prediction. Math Probl Eng; 2021; 2021, pp. 1-10. [DOI: https://dx.doi.org/10.1155/2021/1792201]

13. Saboor, A; Usman, M; Ali, S; Samad, A; Abrar, MF; Ullah, N. A method for improving prediction of human heart disease using machine learning algorithms. Mob Inf Syst; 2022; 2022, pp. 1-9. [DOI: https://dx.doi.org/10.1155/2022/1410169]

14. Sundar, C; Chitradevi, M; Chitradevi, M; Geetharamani, G. Classification of cardiotocogram data using neural network based machine learning technique. Int J Comput Appl; 2012; 47, 14 pp. 19-25. [DOI: https://dx.doi.org/10.5120/7256-0279]

15. Ocak, H; Ertunc, HM. Prediction of fetal state from the cardiotocogram recordings using adaptive neuro-fuzzy inference systems. Neural Comput Appl; 2013; 23, 6 pp. 1583-1589. [DOI: https://dx.doi.org/10.1007/s00521-012-1110-3]

16. Ocak, H. A medical decision support system based on support vector machines and the genetic algorithm for the evaluation of fetal well-being. J Med Syst; 2013; 37, 2 pp. 1-9. [DOI: https://dx.doi.org/10.1007/s10916-012-9913-4]

17. Peterek, T; Gajdoš, P; Dohnálek, P; Krohová, J. Human fetus health classification on cardiotocographic data using random forests. Adv Intell Syst Comput; 2014; [DOI: https://dx.doi.org/10.1007/978-3-319-07773-4_19]

18. Mehbodniya, A; Lazar, AJP; Webber, J; Sharma, DK; Jayagopalan, SKK; Singh, P; Rajan, R; Pandya, S; Sengan, S. Fetal health classification from cardiotocographic data using machine learning. Expert Syst; 2021; [DOI: https://dx.doi.org/10.1111/exsy.12899]

19. Rahmayanti, N; Pradani, H; Pahlawan, M; Vinarti, R. Comparison of machine learning algorithms to classify fetal health using cardiotocogram data. Procedia Comput Sci; 2022; 197, pp. 162-171. [DOI: https://dx.doi.org/10.1016/j.procs.2021.12.130]

20. Alam, MT; Khan, MAI; Dola, NN; Tazin, T; Khan, MM; Albraikan, AA; Almalki, FA. Comparative analysis of different efficient machine learning methods for fetal health classification. Appl Bion Biomech; 2022; 2022, pp. 1-12. [DOI: https://dx.doi.org/10.1155/2022/6321884]

21. The UCI, Machine learning repository, cardiotocography dataset. https://archive.ics.uci.edu/ml/datasets/cardiotocography. Accessed 1 Jan 2023

22. Chawla, NV; Bowyer, KW; Hall, LO; Kegelmeyer, WP. SMOTE: synthetic minority over sampling technique. J Artif Intell Res; 2002; 16, pp. 321-357. [DOI: https://dx.doi.org/10.1613/jair.953]

23. Cao, XH; Stojkovic, I; Obradovic, Z. A robust data scaling algorithm to improve classification accuracies in biomedical data. BMC Bioinform; 2016; 17, 1 pp. 1-10. [DOI: https://dx.doi.org/10.1186/s12859-016-1236-x]

24. Hyperparameter tuning. Grid search and random search. https://www.yourdatateacher.com/2021/05/19/hyperparameter-tuning-grid-search-and-random-search/#:~:text=Grid%20search%20is%20the%20simplest,performance%20metrics%20using%20cross%2Dvalidation. Accessed 1 Jan 2023

25. Bergstra, J; Bengio, Y. Random search for hyper-parameter optimization. J Mach Learn Res; 2012; 13, 1 pp. 281-305.2913701

26. Talukder, MSH; Sarkar, AK. Nutrients deficiency diagnosis of rice crop by weighted average ensemble learning. Smart Agric Technol; 2023; 4, [DOI: https://dx.doi.org/10.1016/j.atech.2022.100155] 100155.

27. Kumari, S; Kumar, D; Mittal, M. An ensemble approach for classification and prediction of diabetes mellitus using soft voting classifier. Int J Cogn Comput Eng; 2021; 2, pp. 40-46. [DOI: https://dx.doi.org/10.1016/j.ijcce.2021.01.001]