High-risk pregnancy prediction using Taguchi-optimized machine learning methods and TOPSIS-based model selection

Abstract

High-risk pregnancies pose a major challenge in maternal and neonatal healthcare, defined as conditions where medical, social, or obstetric factors threaten the health of the mother or fetus. Traditional clinical tools often overlook individual complexities and delay identification, especially in resource-limited settings where mortality rates remain high. This study addresses these gaps by developing machine learning (ML) models for high-risk pregnancy prediction using data from 62 pregnant women at an Iranian hospital (2014–2016). Features were categorized into demographics, pregnancy-related, and complete case (all features). Preprocessing, feature selection, and hyperparameters were optimized using the Taguchi method. Five supervised models, K-Nearest Neighbors (KNN), Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), and Multilayer Perceptron (MLP), were trained and evaluated using multiple performance and statistical criteria. The study contributes an integrated framework combining feature-group comparison, Taguchi-based optimization, and TOPSIS-based multi-criteria model ranking to support balanced model selection. Using real-world hospital-based data from Iran, it jointly predicts major high-risk pregnancy conditions, including IUFD, IUGR, and preterm birth. Results showed pregnancy-related features as stronger predictors, with KNN achieving 88% accuracy and high recall in this category. TOPSIS ranked SVM highest for demographics, KNN for pregnancy-related, and MLP for the complete case. This framework may support clinicians and maternal health centers in earlier identification and prioritization of high-risk pregnancies.

Introduction

High-risk pregnancies represent a critical challenge in maternal and neonatal healthcare, characterized by the presence of medical, social, or obstetric factors that adversely affect the health and well-being of the mother, fetus, or both, leading to elevated risks of mortality, long-term complications, and developmental disorders^1,2. According to the World Health Organization (WHO), a high-risk pregnancy is defined as a condition where the life, health, or welfare of the mother or developing fetus is threatened due to pre-existing medical conditions, pregnancy-related issues, or environmental and social factors³. This definition aligns with clinical perspectives that emphasize the potential for adverse outcomes, such as preterm birth, intrauterine growth restriction (IUGR), intrauterine fetal demise (IUFD), preeclampsia, gestational diabetes, and postpartum complications^4,5. For instance, preterm birth is specifically termed as delivery before 37 weeks of gestation, often triggered by infections, hypertension, placental abnormalities, or prior history, resulting in underdeveloped organs like lungs in the neonate and increased neonatal mortality⁶. Similarly, IUGR occurs when the fetus grows significantly less than expected for gestational age (below the 10th percentile), linked to maternal factors such as chronic diseases, malnutrition, or smoking, and fetal issues, including genetic anomalies, leading to risks like cerebral palsy, stillbirth, and long-term metabolic disorders^7,8. IUFD, defined as fetal death after 20 weeks without signs of life, is associated with chromosomal abnormalities, placental problems, maternal infections, or unexplained causes, carrying profound emotional, psychological, and social impacts on families^9,10.

The importance and urgency of addressing high-risk pregnancies cannot be overstated, particularly in low- and middle-income countries where 95% of the 287,000 annual maternal deaths occur, many of which are preventable through early detection and intervention^3,11. In fragile or crisis-affected regions, maternal mortality rates can be up to five times higher than in stable areas, exacerbated by limited access to quality healthcare³. High-risk pregnancies not only increase maternal morbidity, such as severe preeclampsia, eclampsia, hemorrhage, and mental health issues like perinatal anxiety, but also contribute to neonatal complications, including low birth weight, respiratory distress, and lifelong disabilities^12,13. Early identification is essential for targeted prenatal care, lifestyle modifications, and resource allocation, ultimately improving perinatal outcomes and reducing economic burdens on healthcare systems¹⁴. Traditional assessment methods, reliant on clinical tools and subjective judgments, often fail to account for individual complexities, leading to delayed interventions and suboptimal results^15,16. For example, conventional risk scoring systems may overlook multifaceted interactions in data, resulting in false negatives that heighten the consequences of undetected risks¹⁷.

Several factors influence high-risk pregnancies, broadly categorized into maternal, fetal, and environmental domains¹⁸. Maternal factors include advanced age (> 35 years), obesity (BMI ≥ 30 kg/m² pre-pregnancy), underweight (BMI < 18.5 kg/m²), hypertension, diabetes (pre-gestational or gestational), autoimmune diseases (e.g., systemic lupus erythematosus), smoking, alcohol/substance use, poor prenatal care, low socioeconomic status, and short interpregnancy intervals (< 6 months)^19–21. Fetal factors encompass multiples (twins or more), congenital anomalies, genetic syndromes, and growth discrepancies like IUGR^8,22. Environmental and social influences involve poverty, low education, urban/rural disparities, high-altitude living, and intimate partner violence^23,24. Studies highlight that obesity and smoking synergistically amplify risks: obese pregnant women face higher chances of gestational diabetes, preeclampsia, and cesarean deliveries, while offspring are predisposed to childhood obesity and cardiovascular diseases¹⁹. In India, a national survey revealed a 14.4% prevalence of high-risk pregnancies, driven by adolescent/maternal age extremes, chronic conditions (e.g., asthma, thyroid disorders), and prior adverse outcomes like preterm birth or cesarean Sect²⁰. A meta-analysis of 83 studies further classified multiple high-risk factors into physical (diabetes, hypertension), psychological (depression), social-behavioral (poverty, substance use), and obstetric history (prior miscarriages), noting an upward trend in low-income settings²¹. Autoimmune conditions, such as Sjögren’s syndrome or type 1 diabetes, elevate risks of miscarriage, IUGR, and small-for-gestational-age births²⁵. Gestational weight gain, either insufficient or excessive, correlates with adverse outcomes; underweight women with inadequate gain experience higher rates of preterm premature rupture of membranes, anemia, and small-for-gestational-age infants^26,27. Mapping studies in Indonesia underscore age, parity, prior disease history, and chronic energy deficiency as key predictors, with 75.5% of cases linked to obstetric history²⁸.

Traditional prediction methods for high-risk pregnancies primarily involve clinical guidelines, such as the WHO’s risk categorization or tools like the Obstetric Comorbidity Index, which rely on checklists of known risk factors^3,29. These approaches, while accessible, are limited by their static nature, subjectivity in clinician interpretation, and inability to handle nonlinear interactions in heterogeneous data^15,30. For instance, scoring systems from the 1950s onward often underperform in early prediction of preterm labor or preeclampsia, with sensitivities below 70% in diverse populations^16,31. In contrast, machine learning (ML) methods offer a paradigm shift by analyzing multidimensional datasets, integrating demographics, vital signs, laboratory results, and historical records, to uncover hidden patterns and provide personalized, timely predictions^32,33. Supervised ML algorithms, such as logistic regression, decision trees, random forests, support vector machines, and neural networks, excel in classification tasks for binary outcomes (high-risk vs. low-risk), achieving accuracies often exceeding 85–95%^32,34. For example, ensemble methods like random forests have demonstrated superior performance in predicting hypertensive disorders of pregnancy (HDP) by handling imbalanced data and feature interactions³⁵. Neural networks, particularly multilayer perceptrons (MLPs), integrate vital signs (e.g., blood pressure, glucose levels) to forecast risks with 91% specificity for high-risk cases³². Studies in the Philippines utilized semi-supervised decision trees on sparse data, reaching 97% accuracy by incorporating unlabeled samples³⁶. In fetal heart rate monitoring, ML classifiers distinguish abnormal patterns, correlating probabilities with clinical outcomes like IUGR or distress³⁷. For preterm birth (PTB) prediction, support vector machines and logistic regression, augmented with SMOTE for class imbalance, identified diabetes and obesity as key risks with high sensitivity^38,39. Deep learning hybrids, such as DT-BiLTCN, achieved 98% accuracy on IoT-sourced maternal health data, emphasizing diastolic/systolic pressure and heart rate⁴⁰. Comparative analyses favor gradient boosting (90.64%) and LASSO-regularized models for early PIH detection^34,41. In low-resource settings, ML models like random forests (85%) and explainable AI variants predict PTB in nulliparous women with 94–97% negative predictive value^42,43. Recent advancements include preeclampsia forecasting via routinely collected data (AUC > 0.85) and hybrid ANN-random forest classifiers for risk stratification^44,45. Despite these successes, gaps persist in preprocessing optimization and multi-criteria model selection, particularly in underrepresented populations like Iran, motivating this study’s integration of Taguchi for robust feature engineering and TOPSIS for balanced evaluation^46,47.

The summary of the reviewed studies is presented in Table 1. Many existing studies on high-risk pregnancy prediction primarily focus on developing machine learning models with an emphasis on predictive performance, often prioritizing accuracy-related metrics over methodological robustness. Although some works report additional evaluation criteria, the selection of preprocessing strategies and model configurations is typically empirical and applied uniformly across all cases, without considering differences in data characteristics or problem complexity. Furthermore, systematic optimization techniques and multi-criteria decision-making frameworks are rarely incorporated to guide model selection in a principled manner. These limitations highlight the need for structured approaches that jointly optimize preprocessing steps and objectively rank predictive models, particularly in complex and sensitive clinical prediction tasks such as high-risk pregnancy.

Table 1.

Literature review summary.

Author(s)	Year	Prediction	Data (type/size)	Key features	Models	Taguchi	TOPSIS	Best metric
Moreira et al.39	2019	Preterm birth	App-based	Maternal medical	DT (bagged)/Other
Prema et al.38	2019	Preterm birth	Hospital data	Maternal medical	SVM, LR
Rajbanshi et al.29	2020	High-risk		Demographic, Pregnancy-related	LR
Arabi Belaghi et al.43	2021	Preterm birth	Hospital data	Maternal medical	LR			94–97%
Macrohon et al.36	2022	High-risk	Clinic	Demographic, Pregnancy-related	DT			Acc 97%
Park et al.37	2022	IUGR	Hospital	Pregnancy-related	ML
Raza et al.40	2022	High-risk	IoT data	Maternal medical, Pregnancy-related	Deep hybrid			Acc 98%
Ravi et al.42	2022	High-risk	Clinical	Maternal medical	RF			Acc 85%
Chen et al.41	2023	IUFD		Maternal medical	GB/LASSO-ML			Acc 90.64%
Rahman & Alam34	2023	High-risk	Clinical	Maternal medical	DL
Togunwa et al.45	2023	High-risk	IoT data	Maternal medical	ANN + RF			Acc 95%
Tiruneh et al.44	2024	IUFD	Clinical	Maternal medical	ML			AUC > 0.85
Khadidos et al.31	2024	High-risk	Clinical	Maternal medical	Ensemble ML
Ukrit et al.33	2024	Maternal risk	IoT; n = 1014	Maternal medical	ML
Pi et al.32	2025	High-risk	Clinical data	Pregnancy-related, Maternal medical	NN			Acc 82–91%
Sylvain et al.47	2025	Adverse outcomes	25 hospitals	Maternal medical, Pregnancy-related	RF			Acc 90.6%
This study	—	IUFD / IUGR / Preterm birth	Hospital; n = 62	Pregnancy-related, Demographic	KNN/RF/DT/SVM/MLP	✓	✓	Acc 88% (KNN)

The main contributions of this study are as follows:

• An integrated analytical framework is developed for high-risk pregnancy prediction by combining supervised machine learning models with Taguchi-based optimization and TOPSIS-based multi-criteria decision-making.

• The real-world hospital-based dataset collected in Iran is categorized into three clinically meaningful feature groups, including demographic, pregnancy-related, and complete-case variables, to provide a focused analysis of the predictive contribution of each feature category.

• The Taguchi method is employed to systematically optimize preprocessing procedures, feature selection strategies, and hyperparameter configurations, thereby improving the robustness and efficiency of the prediction framework.

• The TOPSIS method is implemented as a Multi-Criteria Decision-Making tool to rank the performance of the machine learning models based on multiple evaluation metrics for each feature category and the complete-case dataset.

• The proposed framework addresses high-risk pregnancy prediction in a broader clinical context by jointly considering major high-risk pregnancy conditions, including intrauterine fetal demise, intrauterine growth restriction, and preterm birth, rather than focusing on a single adverse pregnancy outcome.

• The proposed framework has potential practical applications in hospitals, prenatal care clinics, maternal health centers, and resource-limited healthcare facilities, where it can support early identification of high-risk pregnancies, closer monitoring, timely referral, and preventive interventions.

Materials and methods

This study is a secondary analysis of an existing clinical dataset from 62 pregnant women at Firoozgar Hospital, Tehran, Iran. The dataset was obtained from the authors of a previously published study⁴⁸, and was used to develop a machine-learning model. No new participant recruitment, intervention, or direct contact with human subjects was performed in the present study. According to the source publication from which the dataset was obtained, the original study was approved by the Ethics Committee of Iran University of Medical Sciences and regional authorities, and written informed consent was obtained from all participants before enrollment, with the right to withdraw at any time. In the present work, only secondary analysis of pre-existing de-identified data was performed.

The research methodology followed a structured flowchart (Fig. 1) outlining the key phases: data collection, comprehensive preprocessing and hyperparameter optimization, model training and evaluation, and multi-criteria ranking. Initially, demographic (e.g., age, education, income, parity) and pregnancy-related features (e.g., history of IUFD, IUGR, gestational hypertension, TPO antibodies, BMI, and common complications) were gathered, with features categorized into three groups: demographics, pregnancy-related, and complete case. Preprocessing steps, including handling missing values (e.g., median, Iterative Imputer, KNN Imputer), standardization (e.g., Standard, Min-Max, Robust), data balancing (e.g., Oversampling, Undersampling, class-weight), and feature selection (e.g., PCA, Filter method, Wrapper method), along with hyperparameter tuning (previously explored via grid or random search methods), were optimized using the Taguchi method to identify robust combinations for each model and category. This integrated optimization minimized sensitivity to noise via signal-to-noise (S/N) ratios and ANOVA analysis, treating the steps as factors with multiple levels (e.g., 3 options per factor for most: missing value imputation methods, standardization techniques, balancing methods, feature selection algorithms, and key hyperparameters like k in KNN or n_estimators in RF). Five ML algorithms, namely K-Nearest Neighbors (KNN), Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), and Multilayer Perceptron (MLP), were then trained and evaluated using a stratified 70/30 train-test split together with 5-fold cross-validation to provide a more stable assessment of model performance under limited-sample conditions. Model performance was assessed using standard metrics: accuracy, precision, recall, specificity, and F1-score, derived from the confusion matrix. Finally, the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) was applied to rank models across categories, prioritizing recall due to its clinical importance in minimizing false negatives. This integrated framework ensured efficient, balanced predictions tailored to high-risk pregnancy detection.

Fig. 1.

Research methodology flowchart: study phases.

Given the real-world clinical nature of the dataset and the limited sample size (n = 62), the study was designed to maximize methodological robustness through a structured optimization and validation strategy. The Taguchi method was employed to systematically optimize preprocessing procedures, feature selection strategies, and hyperparameter configurations while reducing unnecessary experimental variability. In addition, model performance was evaluated using cross-validation to obtain more stable estimates across data partitions. These procedures were adopted to ensure that the reported results reflect consistent performance patterns within the available clinical dataset, while maintaining appropriate caution regarding broader external generalization.

Data collection and preparation

Data were prospectively collected from 62 married pregnant women aged 18–40 years, residing in Tehran, with no history of thyroid disorders, at Firoozgar Hospital between 2014 and 2016. This exclusion of thyroid-related cases was deliberate to focus on non-endocrine influences on pregnancy risks, avoiding confounding factors like hypothyroidism or hyperthyroidism that could independently elevate high-risk status. Inclusion criteria ensured a focused cohort suitable for ML analysis, excluding cases with incomplete records. Features totaled approximately 24, divided into:

• Demographics: Age, parental education and income, occupation, number of pregnancies, successful births after 20 weeks, living children, recent life changes, menstrual history, and recent abortion.

• Pregnancy-related: Recent IUFD, IUGR, amniotic fluid status, gestational hypertension, TPO antibodies, BMI, and common complications.

• Complete case: Complete case integration of all features, representing the full dataset without exclusions beyond initial criteria. Ethical approval was obtained from the hospital’s institutional review board, with informed consent ensuring data anonymity and compliance with privacy regulations (e.g., Iran’s data protection guidelines). The dataset exhibited class imbalance (high-risk vs. low-risk), addressed during preprocessing. Table 2 presents a summary of the features used in this study.

Table 2.

Summary of features used in the study.

Feature Name	Description	Type
Age	Mother’s age in years	Numerical
Mother’s education	Mother’s education level (e.g., high school, Diploma)	Non-numerical
Father’s education	Father’s education level (e.g., high school, Diploma)	Non-numerical
Income	Family income level	Numerical
Mother’s occupation	Mother’s occupation	Non-numerical
Father’s occupation	Father’s occupation	Non-numerical
Number of pregnancies	Total number of previous pregnancies	Numerical
Successful births (> 20 weeks)	Number of full-term births after 20 weeks	Numerical
Living children	Number of surviving children	Numerical
Recent life changes	Presence of recent stressors	Non-numerical
Menstrual history	Irregularities in menstrual cycle (e.g., normal, abnormal)	Non-numerical
Recent abortion	History of recent miscarriage	Numerical
Recent IUFD	History of recent intrauterine fetal death	Numerical
Recent IUGR	History of recent fetal growth restriction	Numerical
Amniotic fluid status	Level of amniotic fluid (e.g., normal, low, high)	Non-numerical
Gestational hypertension	Presence of high blood pressure during pregnancy	Non-numerical
TPO antibodies	Anti-thyroid peroxidase antibodies level	Numerical
BMI	Body Mass Index (kg/m²)	Numerical
Common complications	Common pregnancy complications	Non-numerical
Sex	Fetal sex (e.g., male, female)	Non-numerical
Depression level	Mother’s depression level	Non-numerical
Anxiety level	Mother’s anxiety level	Non-numerical
Death child	History of child death	Non-numerical
High-risk pregnancy	Presence of high-risk pregnancy status	Non-numerical

Data preprocessing

Raw data preprocessing was performed to enhance model robustness and ensure reliable machine learning performance prior to optimization. Given the presence of missing values, mixed data types, and class imbalance in the dataset, multiple preprocessing components were considered, including missing value imputation, feature scaling, class balancing, feature selection, and hyperparameter tuning.

Missing values were handled using three alternative approaches: Median imputation, Iterative Imputer, and KNN Imputer. Feature scaling was performed using Standard (z-score), Min-Max, and Robust normalization techniques. To address class imbalance between high-risk and low-risk pregnancy cases, three balancing strategies were evaluated: Oversampling (e.g., SMOTE), Undersampling, and class-weight adjustment. Feature selection methods included Principal Component Analysis (PCA), filter-based methods (e.g., Chi-square), and wrapper-based methods (e.g., recursive feature elimination). Hyperparameters were tuned using either a grid-search equivalent strategy or a random-search subset strategy over predefined key parameters (e.g., KNN k={3,5,7}, RF n_estimators={50,100,150}).

Hyperparameter tuning was incorporated within the Taguchi optimization framework rather than being treated as a separate post hoc step. In particular, the tuning strategy itself (Grid search versus Random search) was included as one of the candidate factor levels in the Taguchi design, and the final strategy reported for each model and feature category in Table 4 was determined according to the configuration that yielded the most robust performance. The tuned hyperparameters were selected according to the structure of each classifier. Specifically, KNN tuning considered the number of neighbors, weighting scheme, and distance metric; RF tuning included the number of trees and tree-complexity parameters; DT tuning focused on split and depth-related parameters; SVM tuning considered kernel-related and regularization parameters; and MLP tuning included hidden-layer structure, activation function, regularization strength, and learning behavior. The corresponding Taguchi signal-to-noise plots generated in Minitab for each model and feature category are presented in Fig. 2.

Table 4.

Optimal configurations for ML models by feature category.

Category	Model	Missing value	Standardization	Balancing	Feature Selection	Hyperparameter tuning
Demographics	KNN	Iterative imputer	Min-Max	Class-weight	Wrapper method	Grid
	RF	Iterative imputer	Min-Max	Undersampling	Wrapper method	Random search
	DT	KNN imputer	Robust	Class-weight	Filter method	Grid
	SVM	KNN imputer	Standard	Oversampling	PCA	Random search
	MLP	Iterative imputer	Min-Max	Undersampling	Wrapper method	Random search
Pregnancy-related	KNN	Median	Min-Max	Oversampling	Filter method	Grid
	RF	Iterative imputer	Min-Max	Class-weight	PCA	Random search
	DT	Median	Standard	Oversampling	PCA	Grid
	SVM	Median	Standard	Oversampling	Filter method	Random search
	MLP	Median	Robust	Oversampling	Filter method	Random search
Complete case	KNN	Iterative imputer	Min-Max	Undersampling	Wrapper method	Grid
	RF	KNN imputer	Standard	Class-weight	Filter method	Random search
	DT	Iterative imputer	Robust	Undersampling	PCA	Random search
	SVM	Median	Min-Max	Undersampling	Wrapper method	Random search
	MLP	Iterative imputer	Min-Max	Undersampling	Wrapper method	Random search

Fig. 2.

Taguchi optimization plots.

The three feature selection strategies were chosen to represent complementary approaches to dimensionality reduction and variable selection. PCA was included as a projection-based method to reduce feature dimensionality and collinearity by transforming the original variables into a smaller set of components. Filter-based selection was included because it offers a computationally efficient way to rank features according to their statistical relevance to the outcome, independent of the final classifier. Wrapper-based selection was also considered because it evaluates feature subsets in relation to model performance and can therefore capture model-specific predictive utility. Rather than assigning these methods to particular models a priori, all three were treated as candidate preprocessing options within the Taguchi framework, and the final choice for each model and feature group was determined according to the configuration that provided the most robust performance.

Optimization using the Taguchi method

To systematically optimize preprocessing strategies and hyperparameter configurations while reducing computational cost, the Taguchi method was employed. Instead of evaluating all possible combinations exhaustively (e.g., 3⁴ × 2 = 162 for five factors with mixed levels), an L18 orthogonal array design was adopted. Taguchi experiments were conducted separately for each machine learning model and feature category. Each trial was executed in Python (scikit-learn), and performance was estimated using 5-fold cross-validation repeated across five iterations per trial. The resulting accuracy values were then analyzed in Minitab. Accuracy was treated as the quality characteristic and converted to a signal-to-noise (S/N) ratio using the “larger-is-better” criterion to identify robust settings that are least sensitive to noise and variability See table 3.

Table 3.

Factors and levels considered in the Taguchi optimization design.

Factor	Level 1	Level 2	Level 3
Missing value imputation	Median	Iterative Imputer	KNN Imputer
Standardization	Standard (z-score)	Min–Max	Robust
Balancing	Oversampling (e.g., SMOTE)	Undersampling	class-weight
Feature selection	PCA	Filter method	Wrapper method
Hyperparameter tuning	Grid-search	Random-search

Machine learning models

Five supervised classifiers were selected for their suitability in binary classification of high-risk pregnancies (high-risk: 1; low-risk: 0), with the explicit goal of predicting high-risk status to facilitate timely clinical interventions. Each model was chosen based on its proven efficacy in handling medical datasets with mixed feature types and potential imbalances, as detailed below.

K-nearest neighbors (KNN)

The K-Nearest Neighbors model is a supervised and non-parametric learning algorithm used for classification and regression. In this method, to predict the class of a new sample, its distance to existing samples, such as Euclidean distance, is calculated, and the K closest samples are selected. The class of the new sample is determined by the majority vote of its neighbors. KNN, due to its simplicity and lack of assumptions about data distribution, is used in applications such as pattern recognition, medical diagnosis like disease identification, and recommendation systems. However, this model is sensitive to data volume and the selection of K value⁴⁸.

Random forest (RF)

Random Forest is a supervised ensemble-based learning algorithm composed of multiple decision trees. This method creates a large number of decision trees on random subsets of data and features, combining the predictions of each tree through voting for classification or averaging for regression to determine the final outcome. Random Forest, due to its resistance to overfitting and ability to manage high-dimensional data, is used in applications such as fraud detection, financial prediction, and biological data analysis⁴⁹.

Decision tree (DT)

Decision Tree is a supervised learning algorithm that models data using a tree structure of decision rules. Each internal node represents a decision based on a feature, and leaves represent the predicted class or value. This model, due to its high interpretability and simplicity, is applied in classification and regression problems such as risk analysis, disease diagnosis, and customer behavior prediction. However, the decision tree may be sensitive to overfitting unless optimized using techniques such as pruning⁵⁰.

Support vector machine (SVM)

Support Vector Machine is a supervised learning algorithm that separates data in classification or regression problems by finding an optimal hyperplane. This model achieves high accuracy by maximizing the margin between classes and can handle nonlinear problems using the kernel trick. SVM is highly effective in applications such as image recognition, text analysis, email categorization, and medical disease detection due to its ability to manage complex and high-dimensional data⁵¹.

Multilayer perceptron (MLP)

Multilayer Perceptron is a type of artificial neural network that uses multiple layers of neurons to model complex and nonlinear relationships in data. This model includes input, hidden, and output layers trained using algorithms such as backpropagation. MLP, due to its ability to learn complex patterns, is used in applications such as image recognition, natural language processing, and time series prediction. However, it requires precise parameter tuning and high computational resources as limitations⁵².

Performance evaluation

The effectiveness of the machine learning models was evaluated using a confusion matrix, a critical tool for assessing binary classification performance in high-risk pregnancy prediction. The confusion matrix provides a structured summary of model predictions, comprising True Positives (TP: correctly predicted high-risk pregnancies), True Negatives (TN: correctly predicted low-risk pregnancies), False Positives (FP: low-risk pregnancies incorrectly classified as high-risk), and False Negatives (FN: high-risk pregnancies incorrectly classified as low-risk). These elements serve as the foundation for calculating key performance metrics: accuracy, precision, recall, specificity, and F1-score. Each metric offers distinct insights into model performance, enabling a comprehensive evaluation of classification efficacy. In the context of high-risk pregnancy prediction, minimizing false negatives is crucial due to the severe consequences of missing high-risk cases, such as increased maternal or fetal mortality risks. Consequently, recall was prioritized to ensure high sensitivity in detecting high-risk pregnancies, aligning with clinical priorities to reduce undetected risks^53,54.

Accuracy provides a general measure of correct predictions across all classes but can be misleading in imbalanced datasets, where it may overemphasize the majority class (e.g., low-risk pregnancies). Precision quantifies the proportion of predicted high-risk cases that are correct, which is vital when false positives lead to unnecessary medical interventions. Recall, or sensitivity, measures the model’s ability to identify all actual high-risk cases, critical in this study to avoid missing critical cases that could lead to adverse outcomes. Specificity evaluates the correct identification of low-risk cases, ensuring minimal misclassification of healthy pregnancies. The F1-score, as a harmonic mean of precision and recall, provides a balanced measure, particularly useful in imbalanced datasets where both false positives and false negatives carry significant costs^53,55. The formulas for these metrics are as follows:

1

2

3

4

5

In addition to the conventional confusion-matrix-based metrics, balanced accuracy and Matthews correlation coefficient (MCC) were calculated to provide a more reliable evaluation under class-imbalance conditions. Balanced accuracy considers both sensitivity and specificity by calculating their average, and is therefore less affected by unequal class distributions than conventional accuracy⁵⁶. It is defined as follows:

6

Matthews correlation coefficient (MCC) was also used as a balanced measure of binary classification performance. Unlike accuracy, MCC considers all four components of the confusion matrix, including true positives, true negatives, false positives, and false negatives, and is particularly useful when the class distribution is imbalanced^57,58.The MCC value ranges from − 1 to + 1, where + 1 indicates perfect classification, 0 indicates random prediction, and − 1 indicates complete disagreement between predicted and actual labels. MCC is calculated as follows:

7

To further assess threshold-independent discriminative performance, receiver operating characteristic area under the curve (ROC-AUC) was calculated. The ROC curve illustrates the trade-off between the true positive rate (TPR) and the false positive rate (FPR) across different classification thresholds⁵⁴. The ROC-AUC summarizes the overall ability of the model to discriminate between high-risk and low-risk pregnancies across all possible classification thresholds. Higher ROC-AUC values indicate better discriminative performance TPR and FPR are calculated as follows:

8

9

Precision–Recall area under the curve, also referred to as PR-AUC or Average Precision (AP), was additionally calculated because precision–recall analysis is particularly informative for imbalanced datasets, where the positive class is of greater clinical interest^55,59. The precision–recall curve evaluates the relationship between precision and recall across different classification thresholds. where and denote precision and recall at the \:n-th threshold, respectively. Higher AP values indicate better ability to identify the positive class, which in this study corresponds to high-risk pregnancy. Average Precision summarizes this curve as a weighted mean of precision values at different recall levels and is defined as follows:

10

Finally, the Brier score was used to evaluate the quality of predicted probabilities. Unlike threshold-based metrics, the Brier score assesses the agreement between predicted probabilities and actual binary outcomes^60,61. where \:Nis the number of observations, is the predicted probability of the high-risk class for the \:i-th observation, and is the actual binary outcome coded as 0 or 1. Lower Brier score values indicate better calibrated probability estimates. It is calculated as follows:

11

Model performance was evaluated using accuracy, precision, recall, specificity, F1-score, balanced accuracy, and Matthews correlation coefficient (MCC). In addition, receiver operating characteristic (ROC) analysis, precision–recall analysis, and calibration analysis were performed to provide a broader assessment of discriminative and probabilistic performance. To quantify uncertainty, 95% confidence intervals were calculated for the main evaluation metrics. To assess whether model performances differed statistically across classifiers, one-way ANOVA was performed on cross-validation-based performance values for each evaluation metric within each feature group. The ANOVA results were used as supplementary statistical evidence. To maintain consistency with the main reported results, the best-performing model for each metric was identified based on the reported performance values in Tables 6 and 9, while ANOVA was used only to assess whether the differences among models were statistically significant. A p-value below 0.05 was considered statistically significant.

Table 6.

Performance metrics of machine learning models across feature categories with Taguchi optimization.

Category	Model	TP	TN	FP	FN	Accuracy	Precision	Recall	Specificity	F1-Score	Balanced Accuracy	MCC
Demographics	KNN	3	9	1	5	66	75	37	90	50	63	0.329
	RF	5	6	4	3	61	55	62	60	58	61	0.224
	DT	4	8	2	4	66	66	50	80	57	65	0.316
	SVM	6	6	4	2	66	60	75	60	66	67	0.350
	MLP	5	5	5	3	55	50	62	50	55	56	0.125
Pregnancy-related	KNN	7	9	1	1	88	87	87	90	87	88	0.775
	RF	4	9	1	4	72	80	50	90	61	70	0.444
	DT	5	8	2	3	72	71	62	80	66	71	0.433
	SVM	4	9	1	4	72	80	50	90	61	70	0.444
	MLP	5	10	0	3	83	100	62	100	76	81	0.693
Complete case	KNN	4	9	1	4	72	80	50	90	61	70	0.444
	RF	5	7	3	3	66	62	62	70	62	66	0.325
	DT	3	9	1	5	66	75	37	90	50	63	0.329
	SVM	4	9	1	4	72	80	50	90	61	70	0.444
	MLP	5	9	1	3	77	83	62	90	71	76	0.553

Table 9.

ROC-AUC, PR-AUC / average precision, and Brier score across feature groups.

Category	Model	ROC-AUC (high-risk)	ROC-AUC 95% CI	PR-AUC/AP (high-risk)	PR-AUC/AP 95% CI	Brier score
Demographics	KNN	0.794	0.591–0.950	0.666	0.375–0.929	0.210
	RF	0.662	0.370–0.922	0.632	0.339–0.933	0.234
	DT	0.737	0.488–0.946	0.642	0.338–0.936	0.239
	SVM	0.662	0.369–0.931	0.601	0.334–0.944	0.232
	MLP	0.562	0.250–0.857	0.634	0.312–0.895	0.285
Pregnancy-related	KNN	0.813	0.537-1.000	0.757	0.456-1.000	0.151
	RF	0.662	0.363–0.923	0.708	0.405–0.935
	DT	0.750	0.448-1.000	0.811	0.556-1.000	0.206
	SVM	0.275	0.074–0.533	0.377	0.183–0.648	0.277
	MLP	0.850	0.600-1.000	0.882	0.656-1.000	0.158
Complete case	KNN	0.675	0.350–0.975	0.707	0.362–0.970	0.220
	RF	0.644	0.338–0.901	0.624	0.327–0.915	0.242
	DT	0.594	0.323–0.854	0.545	0.276–0.862	0.305
	SVM	0.675	0.369–0.958	0.733	0.430–0.958	0.242
	MLP	0.750	0.463–0.975	0.732	0.415–0.982	0.189

Model ranking with TOPSIS

The Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) was utilized as a multi-criteria decision-making method to rank the five machine learning models (KNN, RF, DT, SVM, MLP) for each feature category (demographics, pregnancy-related, and complete case) based on their performance metrics (accuracy, precision, recall, specificity, and F1-score). TOPSIS constructs a decision matrix with models as rows and evaluation metrics as columns, normalizes the matrix, and assigns weights to each metric such that their sum equals one to reflect clinical priorities in high-risk pregnancy prediction. Weights were assigned based on clinical priorities, with recall receiving the highest importance to reduce false-negative predictions. The method then identifies ideal positive (best) and ideal negative (worst) solutions, computes the Euclidean distance of each model from these ideals, and calculates a closeness coefficient to rank the models, where higher values indicate superior performance. This approach enables a systematic and objective ranking by balancing trade-offs among multiple evaluation criteria. Implemented in Python using NumPy for distance calculations, TOPSIS effectively supports multi-dimensional model evaluation⁶².

Accuracy was selected as the quality characteristic because it provides a comprehensive and stable measure of overall model performance that facilitates consistent comparison across different preprocessing pipelines and model configurations. In the context of this study, accuracy is employed as a methodological evaluation criterion, particularly appropriate for a small clinical dataset where class-specific metrics may be more susceptible to sampling variability. At this stage, accuracy is used to support robust and reproducible configuration assessment rather than to express clinical preference.

In contrast, recall was assigned a higher weight exclusively during the TOPSIS-based model ranking and was not used as the optimization criterion within the Taguchi framework. This distinction reflects the fundamentally different roles of the two methods: Taguchi optimization requires a single performance metric to identify stable and robust configurations, whereas TOPSIS enables multi-criteria evaluation and the explicit incorporation of clinical priorities. Emphasizing recall at the TOPSIS stage ensures that the final model selection prioritizes the reduction of false-negative predictions, which is critical in the clinical context of high-risk pregnancy detection.

Results

Taguchi optimization results

The Taguchi method was employed to optimize preprocessing and hyperparameter configurations for each machine learning model across the three feature categories, including Demographics, Pregnancy-related, and Complete case. The results are summarized in Table 4. Using the L18 orthogonal array design in Minitab, 18 experiments were conducted for each model and feature category, with accuracy used as the quality characteristic under the “larger-is-better” signal-to-noise criterion. Each experiment was evaluated using 5-fold cross-validation to obtain more robust performance estimates. As shown in Table 4, the optimal configurations varied across models and feature groups, indicating that preprocessing choices and tuning strategies should not be assumed to be uniform across all classifiers or data categories. The corresponding signal-to-noise plots are presented in Fig. 2 and illustrate the relative robustness of the selected configurations.

To improve methodological transparency and reproducibility, the principal hyperparameters tuned for each classifier are summarized in Table 5. The selected tuning settings were chosen according to the structure of each model and were explored using either grid search or random search within the Taguchi-based optimization framework. In this way, hyperparameter tuning was treated as part of the integrated preprocessing and configuration selection process rather than as a separate post hoc step.

Table 5.

Model-specific tuned hyperparameters and search strategies used in the optimization framework.

Model	Main tuned hyperparameters	Search strategy used in this study	Representative candidate search space
KNN	n_neighbors; weights; metric	Grid search	n_neighbors = {3, 5, 7, 9, 11}; weights = {uniform, distance}; metric = {euclidean, manhattan, minkowski}
RF	n_estimators; max_depth; min_samples_split; min_samples_leaf; bootstrap	Random search	n_estimators = {50, 100, 200, 300}; max_depth = {None, 10, 20, 30, 40}; min_samples_split = {2, 5, 10}; min_samples_leaf = {1, 2, 4}; bootstrap = {True, False}
DT	criterion; max_depth; min_samples_split; min_samples_leaf	Grid search or random search, depending on feature group	criterion = {gini, entropy}; max_depth = {None, 5, 10, 20, 30}; min_samples_split = {2, 5, 10}; min_samples_leaf = {1, 2, 4}
SVM	kernel; C; gamma (and degree when polynomial kernels were considered)	Random search	kernel = {linear, rbf, poly, sigmoid}; C = {0.1, 1, 10, 100}; gamma = {scale, auto}; degree = {2, 3, 4} when applicable
MLP	hidden_layer_sizes; activation; solver; alpha; learning_rate	Random search	hidden_layer_sizes = {(50,), (100,), (50,50), (100,50)}; activation = {relu, tanh}; solver = {adam}; alpha = {0.0001, 0.001}; learning_rate = {constant, adaptive}

The search strategy shown in Table 5 reflects the type of tuning procedure used to explore the model-specific parameter space, whereas the final optimized configuration for each feature group is reported separately in Table 4; Fig. 2.

Model performance across feature categories

Model performance metrics for the optimized configurations are presented in Table 6 for the Demographics, Pregnancy-related, and Complete case categories. To provide a more comprehensive evaluation of model performance beyond accuracy, precision, recall, and F1-score, Table 6 presents the expanded threshold-based performance results for all models across the three feature groups. In addition to the original classification metrics, confusion matrix-derived counts (TP, TN, FP, and FN), specificity, balanced accuracy, and Matthews correlation coefficient (MCC) are also reported. These additional measures were included to strengthen the robustness of model comparison, particularly in light of the limited sample size and potential class imbalance. The Pregnancy-related category consistently demonstrated superior performance, with the KNN model achieving an accuracy of 88% and high recall (Table 6), indicating strong predictive capability for high-risk pregnancies. In contrast, the Demographics category exhibited lower performance across all metrics, while the Complete case category provided balanced results, leveraging the full feature set (Table 6).

The results in Table 6 indicate that the pregnancy-related feature group provided the strongest threshold-based classification performance among the three feature categories. In this group, KNN achieved the highest overall accuracy and recall, indicating its strong ability to identify high-risk pregnancy cases after Taguchi-based optimization. MLP also showed favorable performance in the pregnancy-related group, particularly in terms of specificity and precision. In contrast, the demographic feature group showed weaker and less stable performance across models, suggesting that demographic variables alone may provide limited predictive information for high-risk pregnancy classification. The complete-case feature group produced intermediate results, with MLP showing the most balanced overall performance among the evaluated classifiers.

To provide a more detailed view of model classification performance, confusion matrices were additionally examined for all five machine learning models across the demographic, pregnancy-related, and complete-case feature groups. These matrices report the distributions of true positives, true negatives, false positives, and false negatives, thereby complementing the threshold-based metrics presented in Table 6. Figure 3 presents the confusion matrix patterns for the demographic, pregnancy-related, and complete-case feature sets, respectively.

Fig. 3.

Confusion matrices of the five machine learning models across the demographic, pregnancy-related, and complete-case feature groups.

To further investigate the impact of the Taguchi optimization, an additional comparison was conducted by evaluating the baseline performance of the machine learning models without applying the Taguchi method. In this setting, models were trained using standard preprocessing and default or commonly used configurations, without systematic optimization of preprocessing steps or hyperparameters. Table 7 presents the accuracy values of different machine learning models across the Demographics, Pregnancy-related, and Complete case feature categories under this non-optimized scenario.

Table 7.

Accuracy of machine learning models across feature categories without Taguchi optimization.

Category	Model	Accuracy (%)
Demographics	KNN	50
	RF	27
	DT	16
	SVM	61
	MLP	50
Pregnancy-related	KNN	72
	RF	55
	DT	55
	SVM	61
	MLP	50
Complete case	KNN	55
	RF	72
	DT	55
	SVM	61
	MLP	61

A direct comparison of Tables 6 and 7 indicates that the application of the Taguchi optimization consistently improves model accuracy across all feature categories. In the absence of Taguchi optimization, accuracy values exhibit greater variability and generally lower performance, particularly for demographic features and certain classifiers. By contrast, Taguchi-optimized models demonstrate higher and more stable accuracy values, with the most notable improvements observed in the pregnancy-related feature category. These results confirm that the Taguchi-based optimization framework enhances the predictive performance of machine learning models across different feature sets.

To further assess the robustness of the reported threshold-based performance metrics, 95% confidence intervals were calculated for accuracy, precision, recall, specificity, F1-score, balanced accuracy, and Matthews correlation coefficient (MCC) across all three feature groups. As shown in Table 8, several intervals remained relatively wide, reflecting the limited sample size and indicating that the reported point estimates should be interpreted with appropriate caution.

Table 8.

95% confidence intervals for threshold-based model performance across feature groups.

Category	Model	Accuracy 95% CI	Precision 95% CI	Recall 95% CI	Specificity 95% CI	F1-score 95% CI	Balanced accuracy 95% CI	MCC 95% CI
Demographics	KNN	44.4–88.9	0.0-100.0	0.0–75.0	66.7–100.0	0.0–80.0	43.8–83.3	-0.175-0.721
	RF	38.9–83.3	20.0-87.5	25.0-100.0	28.6–88.9	25.0-83.3	37.0-83.8	-0.255-0.671
	DT	44.4–88.9	25.0-100.0	14.3–85.7	50.0-100.0	18.2–84.2	42.3–87.5	-0.161-0.756
	SVM	44.4–88.9	28.6–90.0	40.0-100.0	28.6–90.0	36.4–88.9	44.4–88.8	-0.114-0.775
	MLP	33.3–77.8	18.2–81.8	25.0-100.0	18.2–80.0	22.2–80.0	32.5–79.2	-0.350-0.570
Pregnancy-related	KNN	72.2–100.0	60.0-100.0	60.0-100.0	66.7–100.0	66.7–100.0	71.2–100.0	0.433-1.000
	RF	50.0-94.4	33.3–100.0	14.3–85.7	66.7–100.0	22.2–88.9	50.0–90.0	0.000-0.800
	DT	50.0-94.4	33.3–100.0	25.0-100.0	50.0-100.0	30.8–90.9	48.8–92.3	-0.025-0.837
	SVM	50.0-94.4	33.3–100.0	14.3–85.7	66.7–100.0	22.2–88.9	50.0–90.0	0.000-0.800
	MLP	66.7–100.0	100.0-100.0	25.0-100.0	100.0-100.0	40.0-100.0	62.5–100.0	0.395-1.000
Complete case	KNN	50.0-94.4	33.3–100.0	14.3–85.7	66.7–100.0	22.2–88.9	50.0–90.0	0.000-0.800
	RF	44.4–88.9	25.0-100.0	25.0-100.0	40.0-100.0	28.6–87.0	42.5–88.8	-0.158-0.775
	DT	44.4–88.9	0.0-100.0	0.0–75.0	66.7–100.0	0.0–80.0	43.8–83.3	-0.175-0.721
	SVM	50.0-94.4	33.3–100.0	14.3–85.7	66.7–100.0	22.2–88.9	50.0–90.0	0.000-0.800
	MLP	55.6–94.4	50.0-100.0	25.0-100.0	66.7–100.0	33.3–94.1	55.4–95.5	0.120–0.894

Overall, the confidence interval analysis suggests that although several models achieved favorable point estimates, uncertainty remains non-negligible across all feature groups, particularly in the complete-case and demographic analyses.

To facilitate visual comparison of threshold-based robustness across the three feature groups, balanced accuracy and MCC were additionally summarized graphically. Figures 4 and 5 provide comparative views of these measures across all models and feature categories.

Fig. 4.

Balanced accuracy across models and feature groups.

Fig. 5.

Matthews correlation coefficient (MCC) across models and feature groups.

To complement the threshold-based evaluation, additional probability-based performance measures were examined across the three feature groups. As shown in Table 9, ROC-AUC and PR-AUC / Average Precision were used to assess discriminative performance, while Brier score was used to evaluate the quality of predicted probabilities. These measures were included to provide a broader assessment of model behavior, particularly under class imbalance.

The probability-based results in Table 9 provide additional insight into model behavior beyond threshold-based metrics. In the pregnancy-related feature group, MLP achieved the highest ROC-AUC and PR-AUC/AP values, whereas KNN produced the lowest Brier score, indicating better probability calibration.

To complement the threshold-based evaluation, probability-based performance measures were also compared visually across feature groups. Figures 6, 7 and 8 summarize ROC-AUC, PR-AUC / Average Precision, and Brier score, respectively, thereby providing a broader assessment of discrimination and probability quality.

Fig. 6.

ROC-AUC (high-risk class) across models and feature groups.

Fig. 7.

PR-AUC / Average precision (high-risk class) across models and feature groups.

Fig. 8.

Brier score across models and feature groups (lower is better).

To further evaluate discriminative performance, ROC curves were plotted for all five models within each feature group, and the corresponding AUC values are reported in Table 9. These curves provide a visual representation of the trade-off between true positive rate and false positive rate across classification thresholds See fig 9.

Fig. 9.

ROC curves for the five machine learning models across the three feature groups: (a) demographics, (b) pregnancy-related, and (c) complete case.

Overall, the ROC curves were consistent with the AUC values reported in Table 9. The pregnancy-related feature group showed the strongest discriminative performance overall, although the best-performing model varied across probability-based measures.

As shown in Tables 10, 11 and 12, most evaluation metrics did not differ significantly among the five models within each feature group. Significant differences were observed for selected metrics only. In the demographic feature group, significant differences were observed for accuracy, balanced accuracy, ROC-AUC, and Brier score. In the pregnancy-related feature group, significant differences were observed for PR-AUC/AP and Brier score. In the complete-case feature group, a significant difference was observed only for Brier score. For statistically significant metrics, the best-performing model was reported according to the corresponding performance values already presented in Tables 6 and 9.

Table 10.

ANOVA-based statistical comparison of model performance across evaluation metrics for the demographics feature group.

Feature group	Evaluation metric	ANOVA p-value	Significance	Best model if significant
Demographics	Accuracy	0.028	Significant	KNN / DT / SVM
Demographics	Precision	0.071	Not significant	Not interpreted
Demographics	Recall	0.454	Not significant	Not interpreted
Demographics	Specificity	0.383	Not significant	Not interpreted
Demographics	F1-score	0.191	Not significant	Not interpreted
Demographics	Balanced Accuracy	0.041	Significant	SVM
Demographics	MCC	0.059	Not significant	Not interpreted
Demographics	ROC-AUC	0.001	Significant	KNN
Demographics	PR-AUC/AP	0.062	Not significant	Not interpreted
Demographics	Brier score	< 0.001	Significant	KNN

Table 11.

ANOVA-based statistical comparison of model performance across evaluation metrics for the Pregnancy-related feature group.

Feature group	Evaluation metric	ANOVA p-value	Significance	Best model if significant
Pregnancy-related	Accuracy	0.591	Not significant	Not interpreted
Pregnancy-related	Precision	0.187	Not significant	Not interpreted
Pregnancy-related	Recall	0.978	Not significant	Not interpreted
Pregnancy-related	Specificity	0.763	Not significant	Not interpreted
Pregnancy-related	F1-score	0.808	Not significant	Not interpreted
Pregnancy-related	Balanced Accuracy	0.712	Not significant	Not interpreted
Pregnancy-related	MCC	0.560	Not significant	Not interpreted
Pregnancy-related	ROC-AUC	0.649	Not significant	Not interpreted
Pregnancy-related	PR-AUC/AP	0.009	Significant	MLP
Pregnancy-related	Brier score	0.005	Significant	KNN

Table 12.

ANOVA-based statistical comparison of model performance across evaluation metrics for the complete-case feature group.

Feature group	Evaluation metric	ANOVA p-value	Significance	Best model if significant
Complete case	Accuracy	0.441	Not significant	Not interpreted
Complete case	Precision	0.595	Not significant	Not interpreted
Complete case	Recall	0.866	Not significant	Not interpreted
Complete case	Specificity	0.993	Not significant	Not interpreted
Complete case	F1-score	0.741	Not significant	Not interpreted
Complete case	Balanced Accuracy	0.367	Not significant	Not interpreted
Complete case	MCC	0.317	Not significant	Not interpreted
Complete case	ROC-AUC	0.484	Not significant	Not interpreted
Complete case	PR-AUC/AP	0.465	Not significant	Not interpreted
Complete case	Brier score	0.037	Significant	MLP

Correlation results

Figure 10 presents the correlation heatmap of all study variables, including the high-risk pregnancy outcome. The results indicate that maternal mental health indicators exhibited the strongest associations with high-risk pregnancy, with depression showing the highest correlation, followed by anxiety. Clinical and obstetric factors such as gestational hypertension, menstrual history, parity, gravidity, and recent intrauterine growth restriction also demonstrated relatively strong correlations with the outcome. In contrast, anthropometric measures, including weight and body mass index across different time points, as well as socioeconomic variables, showed comparatively weaker associations. The correlation analysis is exploratory in nature, and the identified associations do not establish causal relationships between the predictors and high-risk pregnancy outcomes. Rather, these findings provide an initial overview of variable relationships to inform subsequent model-based analysis.

Fig. 10.

The heatmap illustrates pairwise correlations among study variables, with warmer colors indicating stronger associations and cooler colors representing weaker ones.

In the reviewed literature, the features most frequently highlighted as influential for adverse pregnancy outcomes and high-risk pregnancy prediction are predominantly pregnancy-related and maternal clinical factors. Across prior studies, recurrent themes include obstetric history and pregnancy complications (e.g., hypertensive disorders of pregnancy, fetal growth restriction related indicators, and other pregnancy complications), along with broader maternal health status variables. Demographic characteristics (such as age or socioeconomic proxies) are often included, but they are typically reported as secondary compared with clinical and obstetric variables. Overall, the literature trend suggests that variables directly reflecting the physiological course of pregnancy and maternal medical condition tend to carry the greatest predictive signal, while purely demographic variables contribute less consistently.

In the present study, the correlation heatmap shows a largely similar pattern: the strongest associations with high-risk pregnancy are concentrated in maternal mental health and key obstetric/clinical variables, with depression and anxiety emerging as the most strongly correlated features, followed by pregnancy-related clinical indicators such as gestational hypertension and other obstetric factors (e.g., menstrual history, parity/gravidity, and recent growth-related complications). This partially aligns with prior work in that pregnancy-related and maternal-condition variables remain dominant; however, our results additionally emphasize psychosocial factors (depression and anxiety) as particularly prominent in this dataset. Taken together, the comparison indicates broad agreement with the literature that pregnancy-related and maternal-condition features are generally more informative than demographic variables, while also suggesting that incorporating maternal mental health measures may provide added value for identifying high-risk pregnancies in settings similar to ours.

Best feature category

Comparative analysis of the three feature categories showed that pregnancy-related variables provided the strongest predictive performance for high-risk pregnancy classification. As reported in Table 6, this feature group achieved the most favorable overall threshold-based results, particularly for KNN, which showed high accuracy, recall, and F1-score. These findings suggest that variables such as recent IUFD, IUGR, gestational hypertension, TPO antibodies, BMI, and common complications contain more direct predictive information than demographic variables alone. The complete-case feature set produced moderate performance, but it did not consistently outperform the pregnancy-related feature group.

Model ranking with TOPSIS

The Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) was applied to rank the machine learning models based on five threshold-based performance metrics: accuracy, precision, recall, specificity, and F1-score. The results are presented in Table 13 and visualized in Fig. 11. In the Demographics category, SVM achieved the highest TOPSIS score and ranked first, reflecting the best overall trade-off among the selected evaluation criteria (Table 13; Fig. 11a). In the Pregnancy-related category, KNN ranked first, consistent with its strong accuracy and recall performance (Table 13; Fig. 11b). In the Complete case category, MLP obtained the highest rank, suggesting a more balanced performance when all available features were considered together (Table 13; Fig. 11c). These findings show that TOPSIS provided a structured multi-criteria ranking approach and helped distinguish models based on their overall performance profile rather than a single metric alone.

Table 13.

TOPSIS rankings of machine learning models across feature categories.

Category	Model	TOPSIS score	Rank
Demographics	KNN	0.50	3
	RF	0.44	4
	DT	0.54	2
	SVM	0.60	1
	MLP	0.35	5
Pregnancy-related	KNN	0.797	1
	RF	0.167	4
	DT	0.230	3
	SVM	0.167	4
	MLP	0.548	2
Complete case	KNN	0.61	2
	RF	0.549	3
	DT	0.319	4
	SVM	0.61	2
	MLP	1	1

Fig. 11.

TOPSIS ranking plots for machine learning models across feature categories.

Discussion

This study examined the performance of multiple machine learning models for high-risk pregnancy prediction using demographic, pregnancy-related, and complete-case feature sets within a small real-world clinical dataset. The findings showed that pregnancy-related features provided stronger predictive performance than demographic variables and the complete-case feature set. Among the evaluated models, KNN achieved the most favorable threshold-based performance in the pregnancy-related category after Taguchi-based preprocessing and configuration optimization. These findings suggest that variables directly related to pregnancy history and clinical status may provide more informative signals for identifying high-risk pregnancies than demographic characteristics alone.

The main methodological contribution of this study is the development of a structured prediction framework. By integrating feature-group comparison, Taguchi-based optimization, and TOPSIS-based multi-criteria ranking, the study provides a systematic approach for evaluating machine learning models under limited-data clinical conditions. This is particularly relevant for maternal health datasets, where small sample size, class imbalance, and the need to balance multiple performance metrics can complicate model selection.

The use of hospital-based data from Iran also adds contextual value, as this population and healthcare setting have been less frequently examined in previous studies on maternal health. Unlike studies that focus on a single adverse outcome, the present framework considers high-risk pregnancy more broadly by incorporating major conditions such as intrauterine fetal demise, intrauterine growth restriction, and preterm birth. From a clinical perspective, such a framework may support future decision-support systems by helping clinicians identify high-risk pregnancies earlier and prioritize monitoring, referral, and preventive care.

The predominance of pregnancy-related and clinical features identified in this study is consistent with prior research highlighting obstetric history and medical conditions as primary determinants of high-risk pregnancy. Kuppusamy et al. reported that factors such as gestational hypertension and prior adverse outcomes were key contributors to high-risk pregnancy classification²⁰. Similarly, Zhang et al. identified hypertension and diabetes as significant predictors in their meta-analysis, reinforcing the importance of clinical variables in predictive models²¹. Although some studies have reported higher accuracy using larger datasets or deep learning architectures^32,40, the performance achieved in the present study is comparable given the limited sample size. Moreover, unlike many previous works that rely mainly on single-metric evaluation, this study applies a multi-criteria ranking approach, providing a more balanced comparison of model performance⁴².

The strong performance of the KNN model, particularly in the pregnancy-related feature set, can be attributed to its ability to capture local patterns in small and heterogeneous datasets without imposing strict distributional assumptions. This characteristic makes KNN well suited for clinical data with mixed feature types and nonlinear relationships³². In addition, correlation analysis revealed that maternal mental health indicators, especially depression and anxiety, exhibited the strongest associations with high-risk pregnancy outcomes. These findings are consistent with previous studies emphasizing the contribution of psychosocial and mental health factors to adverse pregnancy outcomes and maternal complications^16,25, highlighting the importance of incorporating mental health variables alongside traditional obstetric features.

From a clinical perspective, prioritizing recall in model evaluation is critical, as false-negative predictions may result in missed identification of high-risk pregnancies and severe maternal or fetal consequences^11,16. The proposed framework demonstrates the potential of integrating optimized machine learning models into early screening and clinical decision-support systems, particularly in resource-limited settings. At the policy level, the findings underscore the importance of incorporating maternal mental health assessment into routine prenatal care and data collection strategies, aligning with global recommendations for proactive maternal health interventions^3,11. The use of relatively simple and interpretable models, such as KNN, further supports feasibility and scalability for real-world clinical implementation.

The additional confidence intervals and supplementary statistical analyses provide a more transparent interpretation of the reported results. Although the sample size was limited, the use of cross-validation, Taguchi-based optimization, and multi-metric evaluation helped reduce the dependence of the findings on a single performance measure or a single modeling configuration. The results, therefore, provide consistent evidence regarding the relative usefulness of pregnancy-related features and the performance patterns of the evaluated models within the available clinical dataset. At the same time, external validation using larger, multi-center datasets would further strengthen the generalizability of these findings and support future clinical translation.

Although recent deep learning, transformer-based, and hybrid architectures have shown strong performance in many prediction tasks, their application in the present study was constrained by the small size and tabular structure of the available hospital-based dataset. Under such conditions, highly complex models may be more vulnerable to overfitting and may not provide reliable comparisons without larger cohorts and external validation. Therefore, this study focused on representative and well-established machine learning classifiers to evaluate the effectiveness of the proposed Taguchi-TOPSIS framework. Future studies with larger, multicenter, and multimodal maternal health datasets should further investigate whether more advanced architectures can improve predictive performance and clinical generalizability.

From a clinical application perspective, the proposed framework could serve as a supportive decision-assistance tool in hospitals, prenatal clinics, maternal health centers, and resource-limited healthcare facilities. Routinely collected demographic, obstetric, clinical, and psychosocial variables could be entered during antenatal visits to generate an early high-risk pregnancy flag. This output should not replace clinical judgment; rather, it may help clinicians prioritize closer assessment, timely referral, and individualized follow-up for women at higher risk. Such an application is consistent with antenatal care recommendations emphasizing early risk identification, detection of pregnancy-related complications, and referral to appropriate levels of care^3,11.

Successful implementation of such a framework in real-world healthcare systems requires careful consideration. Potential challenges include incomplete or inconsistent clinical data, integration with existing medical records or digital workflows, staff training, model interpretability, external validation, and adaptation to local healthcare infrastructure. Therefore, the proposed framework should be viewed as a supportive screening and prioritization tool, and future studies should evaluate its performance and usability in larger, external, and multi-center clinical settings¹⁵.

Limitations and future directions

Despite the structured methodology, this study has several limitations. The small sample size and single-center design at Firoozgar Hospital in Tehran limit the generalizability of the findings to broader populations and healthcare settings. In addition, the exclusion of women with thyroid disorders, while useful for controlling endocrine-related confounding, may restrict the applicability of the framework to pregnancies affected by thyroid dysfunction. The absence of genetic, biochemical, and longitudinal follow-up data also limits the ability of the models to capture more complex biological and temporal risk patterns. To partially address these constraints, the study incorporated cross-validation, Taguchi-based optimization, confidence interval analysis, and supplementary statistical testing.

Future research should extend this work in several directions. First, larger datasets from diverse geographical, demographic, and clinical populations are needed to evaluate the external validity of the proposed framework. Second, incorporating additional clinical, laboratory, genetic, psychosocial, and longitudinal maternal health variables may improve risk characterization and predictive accuracy. Third, comparative studies between machine learning models and traditional clinical risk assessment tools are required to clarify the added value of data-driven approaches in routine antenatal care. Finally, more advanced machine learning and deep learning architectures should be explored when sufficiently large and externally validated datasets are available, particularly in studies using multimodal maternal health data.

Conclusion

This study demonstrated the efficacy of machine learning in predicting high-risk pregnancies, achieving 88% accuracy with the KNN model in the Pregnancy-related category. The Taguchi-TOPSIS framework optimized preprocessing and model selection, highlighting pregnancy-related features as key predictors. These findings offer a scalable approach for early risk detection in resource-limited settings like Iran, contributing to reduced maternal mortality and improved prenatal care. Future translation of this framework into clinical practice will require external validation, workflow integration, and adaptation to local healthcare infrastructure. Overall, the main contribution of this study is the development of a structured and clinically applicable machine-learning evaluation framework, rather than the proposal of a new standalone algorithm. Future work should focus on larger datasets and genetic factors to enhance model robustness and applicability.

Author contributions

Maryam Mousavi Nogholi: Conceptualization, Study design, Data collection, Data analysis, Writing – original draft.Ashkan Mozdgir: Conceptualization, Study design, Supervision, Writing – review & editing.Younes Javid: Conceptualization, Writing – review & editing.All authors have read and approved the final manuscript.

Data availability

The datasets analyzed during the current study can be obtained from the first author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

This study is a secondary analysis of an existing dataset obtained from the authors of a previously published study entitled “Evaluation of the relationship between serum anti TPO antibody levels and mood disorders in euthyroid pregnant women.” According to the source publication, the original study was approved by the Ethics Committee of Iran University of Medical Sciences and regional authorities. Written informed consent was obtained from all participants before enrollment, and participants had the right to withdraw from the study at any time.