Development and interpretability of a machine learning model for postoperative hypothermia prediction in placenta previa Cesarean section: A retrospective cohort study

Abstract

Objective

To develop and validate an explainable machine learning prediction model for perioperative hypothermia in patients undergoing cesarean delivery for placenta previa, and to identify key predictive factors to support clinical individualized temperature management and nursing decision-making.

Methods

This study is a retrospective secondary analysis conducted using BioStudies public databases. Ten machine learning algorithms, including logistic regression, decision tree, and random forest, were used to build hypothermia prediction models. Accuracy, precision, recall, F1-score, and AUC were selected as evaluation metrics to screen for the optimal model, and LIME was applied to generate individualized explanations for single-sample prediction outputs. A total of 167 real-world patients served as the independent test cohort, while 1000 synthetic data cases generated by the Synthetic Data Vault (SDV) algorithm served as the training cohort.

Results

Performance comparisons across models showed the random forest model reached an AUC of 0.781 in the training set, but only 0.638 in the test set, a sign of severe overfitting. Gradient boosting models had a universally low test set recall of below 0.25, carrying a high risk of missed diagnosis. The logistic regression (LR) model exhibited the optimal overall performance: it attained the highest test-cohort accuracy among all evaluated models, with an AUC of 0.705, F1-score of 0.430, and recall of 0.327. Feature importance analysis found intraoperative blood loss to be the strongest predictor of hypothermia, followed by body weight as the second core contributing factor. LIME can intuitively visualize feature contributions for individual sample predictions. These outputs enable nurses to develop tailored warming intervention regimens.

Conclusion

The logistic regression prediction model developed in this study may serve as a clinically useful tool with adequate generalization capacity and clinical interpretability for identifying hypothermia risk in patients undergoing cesarean delivery for placenta previa.

Introduction

Perioperative hypothermia is among the most common complications of surgical procedures.^1,2 In patients undergoing cesarean delivery, hypothermia can also impair fetal thermoregulation via placental circulation, elevating the risks of neonatal acidosis and low Apgar scores.^3,4 Studies have shown that the incidence of hypothermia in high-risk obstetric surgical populations can exceed 60%.^4,5 Accumulating evidence confirms that intraoperative hypothermia increases the risk of adverse events including elevated intraoperative blood loss and postoperative shivering.^4–8 Placenta previa is a common type of high-risk pregnancy in obstetrics. The average intraoperative blood loss in affected patients can reach 1000–2000 mL, leading to a high incidence of hypothermia secondary to massive blood loss.^9,10 Hypothermia further exacerbates coagulation dysfunction and leads to increased postpartum hemorrhage, making it a key challenge in the perioperative management of placenta previa. ¹¹ However, current clinical temperature management for patients with placenta previa still relies on empirical judgment, making it difficult to deliver precise interventions.

Most existing prediction models for perioperative hypothermia are developed for populations undergoing general surgery or general cesarean delivery. They do not incorporate the unique high-risk features of patients with placenta previa, resulting in poor predictive performance in this specific group.^12–14 Recent studies have demonstrated the successful application of machine learning across diverse clinical domains, including prediction of postoperative complications following general surgery, ¹⁵ emergency cesarean section prediction, ¹⁶ postpartum hemorrhage risk assessment, ¹⁷ preterm birth prediction, ¹⁸ neonatal intensive care unit admission prediction, ¹⁹ and cancer recurrence prognosis. ²⁰ These studies collectively underscore the growing utility of machine learning in obstetric and perioperative care. Machine learning technology can achieve accurate risk prediction by mining underlying associations in clinical data, but the “black-box” nature of traditional machine learning models prevents clinicians and nurses from understanding their decision-making logic.^21–23 Synthetic data generation technology can expand the training sample size without compromising patient privacy. Explainable artificial intelligence (XAI) addresses the “black-box” limitation of models by intuitively demonstrating the contribution of each feature to the prediction outcome. The combination of these two technologies provides a new solution for perioperative risk prediction and nursing intervention guidance in high-risk populations.^24,25

Therefore, this study focuses on patients undergoing cesarean delivery for placenta previa, a high-risk population for hypothermia. We develop prediction models using a fusion approach of real-world data and SDV-generated synthetic data, and screen for the model with optimal generalization ability through multi-algorithm comparison. Unlike traditional statistical approaches that rely on pre-specified linear assumptions, machine learning algorithms can capture complex nonlinear interactions among predictors without requiring a priori specification of variable relationships. Furthermore, the integration of XAI techniques ensures clinical transparency. Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines to ensure methodological rigor. ²⁶ This study aims to provide a clinically practical decision-making tool for perioperative temperature management and nursing interventions in patients with placenta previa.

Materials and methods

Study design and data source

This retrospective analysis utilized the publicly accessible BioStudies database, an open research repository containing fully de-identified and anonymized participant data.

Study subjects were pregnant patients with placenta previa who underwent elective cesarean delivery under general anesthesia between May 2019 and 2023, retrieved from the publicly available biostudies database. A total of 167 patients were finally included as the real-world test cohort.

Synthetic data generation

The Synthetic Data Vault (SDV) algorithm was used to build the generative model. This algorithm learns variable distributions, inter-variable correlations, and temporal characteristics from real-world data to generate synthetic data that matches the distribution of real data but contains no patient identifiable information. It complies with medical research ethics requirements and prevents patient privacy leakage. In this study, 167 cases of real-world data were used as the training input for the SDV model, with the generated sample size set to 1000 cases. Homogeneity score was used to evaluate the distribution consistency of categorical variables, and the homogeneity between synthetic data and real data was verified to ensure that the synthetic data conformed to clinical logic and clinical characteristics.

Variable definition

A total of 13 predictor variables and 1 outcome variable were included in this study. All variables are routinely collected indicators from clinical nursing assessments and anesthesia records: 1) Predictor variables: Group, Age, Height, Weight, Body mass index, Gestational age at delivery, Prior cesarean delivery, Hypothyroidism, Diabetes mellitus, Preoperative temperature, Intraoperative fluid, Intraoperative bleeding, Operation duration; 2) Outcome variable: Immediate postoperative hypothermia, defined as tympanic membrane temperature <36°C measured by nursing staff immediately after the end of surgery.

Model construction and data preprocessing

The 1000 synthetic data cases served as the training cohort, and the 167 real-world data cases were used as the independent external test cohort. The training set and test set were completely independent to avoid data leakage. This split strategy-using independently sourced synthetic data for training and real-world single-center data for testing-represents a form of temporal and distributional separation that mitigates overfitting risk while acknowledging the limitations of single-center validation. No formal feature selection algorithm was applied; all clinically relevant predictor variables were retained based on clinical reasoning and prior literature, as the sample size did not support data-driven dimensionality reduction.

Data preprocessing steps were as follows: Categorical variables were converted to numerical variables using one-hot encoding; continuous variables were standardized using Z-score standardization. Regarding missing data handling, variables with >30% missingness were excluded from the analysis. For remaining variables with missing values, mode imputation was applied for categorical variables and median imputation for continuous variables.The training cohort used 5-fold cross-validation to optimize model hyperparameters. For each fold, 80% of samples were randomly divided into the training subset and 20% into the validation subset to ensure stable model performance. Hyperparameter tuning was performed exclusively within the training cohort using grid search combined with manual tuning; no information from the test set was used at any stage of model development or hyperparameter optimization, thereby preventing data leakage.

A total of 10 commonly used machine learning models were constructed in this study, covering different types including linear models, tree-based models, and instance-based models, to comprehensively screen the optimal model: 1) Linear models: Logistic Regression, Linear Support Vector Machine (Linear SVC); 2) Tree-based models: Decision Tree, Random Forest Classifier, Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM, LGBM), CatBoost, Adaptive Boosting (AdaBoost); 3) Other models: k-Nearest Neighbors (kNN). Hyperparameter optimization of all models was performed using grid search and manual tuning.

Model evaluation and interpretability analysis

Five core indicators were used to evaluate model performance, taking into account clinical needs especially application requirements in nursing scenarios: 1) Accuracy; 2) Precision; 3) Recall; 4) F1-score; 5) Area Under the Receiver Operating Characteristic Curve (AUC). AUC values were interpreted using the following benchmarks: 0.5 = no discrimination, 0.5–0.7 = poor discrimination, 0.7–0.8 = acceptable discrimination, 0.8–0.9 = excellent discrimination, >0.9 = outstanding discrimination. Confusion matrices were generated for the optimal model to visualize classification performance.

The LIME algorithm was used to perform interpretability analysis of the optimal model. For the prediction result of an individual patient, LIME can quantify the direction and magnitude of contribution of each feature to the prediction result, which is intuitively displayed in the form of a waterfall plot.

Statistical methods

Normally distributed continuous variables are described as mean ± standard deviation, and between-group differences were compared using the two independent samples t-test. Non-normally distributed continuous variables are expressed as median (interquartile range) [M (Q1, Q3)], and between-group comparisons were performed using the Mann-Whitney U test. Categorical variables are described as frequency (percentage), and between-group comparisons were performed using the chi-square test or Fisher's exact test. Linear correlation between variables was analyzed using Pearson product-moment correlation, with P < 0.05 set as the significance threshold for statistical significance. All statistical analyses were performed using R software.

Results

Validation of synthetic data effectiveness

Baseline characteristics of the synthetic data cohort (training set) and the hospital population cohort (test set) are compared in Table 1. The synthetic data cohort had a homogeneity score of 91.3%, which was higher than 80%, confirming that the synthetic dataset was eligible for machine learning model construction (Table 1).

Table 1.

Baseline clinical characteristics of patients.

	Training	Testing	P-value
Numbers	1000	167
Age (years)	35.00 (32.00–38.00)	35.00 (32.00–38.00)	0.808
Height (cm)	161.74 (5.57)	161.79 (5.67)	0.909
Weight (kg)	68.87 (61.93–77.26)	68.50 (62.35–76.00)	0.766
Body mass index (kg/m2)	26.60 (24.20–29.00)	26.40 (23.90–29.10)	0.519
Gestational age at delivery (weeks)	36.00 (35.00–37.00)	36.00 (35.00–37.00)	0.566
Preoperative temperature (°C)	36.80 (36.70–37.00)	36.80 (36.70–37.00)	0.925
Intraoperative fluid (ml)	1392.50 (861.25–2054.75)	1300.00 (1075.00–1800.00)	0.975
Intraoperative bleeding (ml)	1211.00 (863.00–1689.25)	850.00 (712.50–1100.00)	<0.001
Operation duration (min)	91.00 (73.00–110.00)	90.00 (75.00–105.00)	0.718
Group (Hybrid operating rooms)			0.020
No	425 (42.50%)	55 (32.93%)
Yes	575 (57.50%)	112 (67.07%)
Prior cesarean delivery			<0.001
0	561 (56.10%)	122 (73.05%)
1	418 (41.80%)	37 (22.16%)
2	21 (2.10%)	7 (4.19%)
3	0 (0.00%)	1 (0.60%)
Hypothyroidism			0.401
No	955 (95.50%)	157 (94.01%)
Yes	45 (4.50%)	10 (5.99%)
Diabetes mellitus			0.164
No	818 (81.80%)	144 (86.23%)
Yes	182 (18.20%)	23 (13.77%)
Hypothermia			0.003
No	564 (56.40%)	115 (68.86%)
Yes	436 (43.60%)	52 (31.14%)

Variable correlation analysis

Pearson correlation analysis between variables showed that the top three positive correlation coefficients associated with hypothermia incidence were for intraoperative blood loss and hypothermia, operation duration and hypothermia, and intraoperative fluid volume and hypothermia, respectively. The highest negative correlation coefficient was observed between body weight and hypothermia (Figure 1).

Figure 1.

Heatmap of correlations among clinical variables.

Performance comparison of different machine learning models

The results are as follows: The Random Forest model delivered the best training set performance, with an AUC of 0.781 and F1-score of 0.619, but its test set AUC was only 0.638 and F1-score was 0.376. Gradient boosting models (GBDT, XGBoost, LGBM, CatBoost) achieved a training set precision of above 0.70, but their recall was generally lower than 0.30, and all had a test set recall below 0.25. Traditional linear models (logistic regression, linear SVC) had the optimal generalization ability, with performance deviations between training and test sets all <10% and no risk of overfitting (Figure 2).

Figure 2.

Area under the receiver operating characteristic (ROC) curve (AUC) results of each algorithm in the test cohort.

Among all models, the logistic regression model had the best comprehensive performance: its test set accuracy was 0.731 (highest among all models), F1-score was 0.430 (highest among all models), recall was 0.327 (highest among all models), and AUC reached 0.705, close to the highest AUC of 0.712 achieved by linear SVC. It also balanced precision and recall simultaneously, and was therefore selected as the final optimal model.

Model feature importance analysis

Results of feature importance ranking for the LGBM model showed that: Intraoperative blood loss was the strongest predictive feature for hypothermia, and body weight was the second core predictive factor (Figure 3).

Figure 3.

Model feature importance analysis.

LIME single-sample interpretability results

A typical patient was selected for LIME interpretation. The model predicted a 38% probability of hypothermia occurrence for this patient, with a final classification of low hypothermia risk, though a certain degree of occurrence risk remained. The LIME waterfall plot demonstrated the logic of feature contributions: Intraoperative blood loss was the strongest risk factor, and operation duration was the second largest risk factor (Figure 4).

Figure 4.

LIME single-sample interpretability results.

Discussion

This is the first study to develop a perioperative hypothermia prediction model for patients undergoing cesarean delivery for placenta previa, a high-risk population for hypothermia, using a multi-source data fusion approach. We screened for the logistic regression model with optimal generalization ability via multi-algorithm comparison, and implemented individualized risk interpretation combined with LIME explainability analysis. All predictor variables are routinely collected indicators in clinical nursing practice, which impose no additional medical burden. The model may be directly embedded into obstetric nursing clinical decision support systems, providing an implementable decision-making tool for perioperative temperature management and nursing interventions.

Regarding internal validity, the use of 5-fold cross-validation within the training cohort, combined with strict separation of training and test sets, helps mitigate overfitting and data leakage. The logistic regression model's train-test performance gap of <5% further supports internal consistency. However, the use of synthetic data for training introduces a potential limitation: while the SDV algorithm preserves marginal distributions and pairwise correlations, it may not capture higher-order interactions present in real clinical data, potentially limiting the ecological validity of the learned patterns.

Regarding external validity, the independent real-world test cohort provides preliminary evidence of generalizability. Nevertheless, the single-center design and relatively small test sample constrain the strength of external validity claims. Multi-center validation in diverse clinical settings remains essential before clinical deployment.

Application value of synthetic data in small-sample clinical and nursing research

Small sample size is a common limitation of single-center clinical studies, especially those focusing on rare high-risk pregnancies. Direct use of small datasets for machine learning model training easily leads to overfitting, resulting in poor generalization ability and limited clinical applicability. Meanwhile, the high requirements for privacy protection of obstetric patients create ethical and policy barriers to multi-center data sharing, which further restricts the implementation of large-sample studies.

In this study, synthetic data were generated using the SDV algorithm. No statistically significant differences were observed in baseline characteristics between synthetic and real data, which effectively expanded the training sample size while eliminating the risk of patient privacy leakage. Previous studies have confirmed that anesthesia depth prediction models trained on synthetic data have equivalent performance to those trained on real data, and can effectively address the privacy concerns associated with multi-center data sharing. ²⁷

Clinical and nursing significance of core predictive factors

This study showed that intraoperative blood loss is the strongest predictor of hypothermia in patients with placenta previa, which aligns closely with the surgical characteristics of placenta previa. In patients with placenta previa, the placenta is attached to the lower uterine segment, with abundant sinusoids at the intraoperative dissection surface. Intraoperative blood loss can reach 3 to 5 times that of uncomplicated cesarean delivery. Massive blood loss carries away a large amount of body heat, while infusion of room-temperature fluids and blood products further reduces core body temperature.^28,29 This finding reminds nursing staff to pay close attention to changes in blood loss during surgery for patients with placenta previa.

Body weight was identified as the second core predictive factor in this study. Patients with lower body weight have less fat reserve and faster heat loss, leading to a higher risk of hypothermia, which is consistent with findings from previous studies in general surgery populations. ³⁰ Low body weight may also be associated with lower blood volume, which could independently contribute to hypothermia through reduced thermoregulatory capacity. In our dataset, the correlation between body weight and intraoperative blood loss was modest, suggesting that body weight may exert its effect through both direct (reduced insulation) and indirect (potentially lower circulating volume) pathways. Future studies incorporating direct hemoglobin or hematocrit measurements may help disentangle these mechanisms. Prior cesarean delivery history was identified as an independent risk factor in this study, which fills a gap in existing research.

Value of explainability for clinical and nursing promotion

The “black-box” nature of machine learning models is a major barrier limiting their clinical application. When nursing staff cannot understand the decision-making logic of a model, it is difficult for them to trust and adopt its prediction results. ³¹ In this study, the LIME algorithm was used to achieve individualized interpretation of single-sample prediction results, which can intuitively demonstrate the direction and weight of contribution of each feature to the patient's risk. It helps nursing staff not only know whether a patient is at high risk, but also understand why the patient is at high risk, so that targeted intervention measures can be developed to truly realize individualized perioperative temperature nursing.

Study limitations and prospects

This study has certain limitations that warrant careful consideration:

1) The real-world test cohort was from a single center with a small sample size. This limits statistical power and the precision of performance estimates. In the future, multi-center large-sample data should be included to further verify the generalization ability of the model, especially its applicability to populations in different regions and hospitals of different levels. Calibration analysis and decision curve analysis should be incorporated in future validation studies to more comprehensively assess clinical utility; 2) Variables that may affect body temperature, such as operating room temperature, anesthetic drug dosage (including type and concentration of volatile anesthetics, opioid equivalents), duration of anesthesia, neonatal weight, and timing of warming measures implemented by nursing staff, were not included in this study. We acknowledge that these are important confounding variables for intraoperative body temperature regulation. Their absence represents a significant limitation of the current model, as they may partially account for unexplained variance in hypothermia outcomes. Specifically, operating room temperature and active warming measures are well-established determinants of perioperative thermoregulation, and their omission means the model may underestimate or overestimate risk in settings with different thermal environments. Future prospective studies should systematically capture these variables to build more comprehensive prediction models; 3) This study is a retrospective study. Prospective clinical intervention studies are needed in the future to verify whether model-based individualized warming nursing measures can actually reduce the incidence of hypothermia and improve maternal and neonatal prognosis in patients with placenta previa; 4) The use of synthetic data for training, while necessary given the small sample size, introduces a distributional mismatch between training and test data that warrants further investigation. The model's performance in this study may represent a conservative estimate, as the synthetic-real data gap likely attenuates rather than inflates performance metrics; 5) The moderate AUC (0.705) indicates acceptable but not excellent discrimination, suggesting the model may be most useful as part of a multimodal assessment rather than a standalone diagnostic tool.

Despite these limitations, the model's reliance on routinely collected clinical variables, combined with its interpretability and transparent risk factor identification, provides a practical foundation for clinical implementation. The pathway to utility may involve: (a) integration into electronic health record systems for real-time risk flagging, (b) prospective validation to establish clinical effectiveness, and (c) incorporation of additional perioperative variables to improve predictive accuracy.

Conclusion

Based on real-world data of patients undergoing cesarean delivery for placenta previa, this study used SDV-generated synthetic data to expand the training sample. The constructed logistic regression prediction model may serve as a clinically useful tool with good generalization ability and clinical interpretability for identifying patients at risk of perioperative hypothermia. Key predictive factors, including intraoperative blood loss, body weight, and prior cesarean delivery history, may guide nursing staff in developing individualized warming strategies. Prospective multi-center validation is warranted before clinical implementation.