Abstract
Objective
Preterm birth (PTB), defined as delivery before 37 weeks of gestation, is the leading cause of perinatal morbidity and mortality. Timely prediction of PTB is crucial for physicians to take preventive action. In this work, we propose an AI-based clinical decision support system to address this challenge.
Methods
We use an existing dataset containing demographic and clinical variables from 973 pregnant women and begin by examining it for potential data-level bias and representativeness issues. Upon identifying an outcome prevalence imbalance in one of the variables, we revised the dataset through a preprocessing adjustment. We then applied several Machine Learning (ML) algorithms to predict PTB as a binary outcome.
Results
During internal model evaluation we found that an optimized ensemble (voting) of Logistic Regression and XGBoost performed best, achieving an accuracy of 96% and a recall of 98%. We also compared the predictive performance using the original (biased) and revised (de-biased) datasets, observing statistically significant improvements in the revised dataset as confirmed by the Wilcoxon signed-rank test, with p-values less than 0.05 for Accuracy, Recall, and F1-score.
Conclusion
The results highlight the importance of data quality and dataset representativeness in developing reliable and trustworthy AI-based applications for PTB prediction.
Keywords
Introduction
Preterm birth (PTB) is defined as delivery before 37 weeks of gestation and is the leading cause of perinatal morbidity and mortality in developed countries. The frequency of PTBs is about 12–13% in the USA and 5–9% in many other developed countries. 1 Thus, prediction and prevention of PTB is a critical focus in obstetric care. Apart from established clinical practice, 2 advances in Artificial Intelligence (AI) 3 and in particular machine learning (ML), have created powerful tools to assist physicians in predicting PTB.4,5
Several studies have demonstrated the potential of ML models to predict PTB with high accuracy; the initial work by Catley et al. 6 used artificial neural networks (ANNs) to predict PTB in low-risk pregnancies increasing the network’s sensitivity to 54.8%, compared to just over 20% for the non-artificially distributed PTB model. Similarly, Andrade et al. 7 apply ANNs on a dataset of 524 singleton pregnancies to predict PTB using clinical and ultrasonographic variables achieving sensitivity of approximately 33%. Moreover, Olatunji Idowu et al. 8 applied Random Forest in a dataset containing 262 records for women who delivered at term and 38 who delivered prematurely achieving a sensitivity of 97%. Similarly, Kloska et al. 9 applied a linear SVM with boosted parameters on a small dataset of 50 women achieving an accuracy of 82%.
In the context of the increasing adoption of AI and ML in health-related applications, one major barrier is the potential lack of trust of the human (i.e., physician or other health care professional) towards these applications. 10 Specifically, when building predictive models, the quality of the results is highly related to the volume and quality of the data with which the ML algorithms are trained and validated. Datasets that contain Inadequate, biased, or non-representative data can compromise accuracy and limit generalizability, in this way reducing clinicians’ confidence. Additionally, the black-box nature of many ML models makes it difficult for professionals to understand or justify predictions, raising ethical and legal concerns. Enhancing transparency, improving data quality, and ensuring rigorous validation are therefore essential for fostering trust and wider adoption in healthcare. Despite the growing body of literature on PTB prediction, most studies primarily focus on algorithmic performance, while significantly less attention has been given to the effect of dataset bias and data representativeness on the reliability of the resulting predictions.
The aim of this study is to develop a fair, bias-free, AI-based clinical decision support system for the timely prediction of preterm birth (PTB) and to evaluate its performance across different machine learning algorithms. The main contribution of this work is the systematic investigation of bias and representativeness issues in the clinical dataset, and the assessment of how bias mitigation strategies such as correcting outcome prevalence imbalance affects predictive performance. Specifically, the study seeks to ih44nvestigate the presence of bias in a dataset containing clinical, demographic and pregnancy variables, assess the impact of bias mitigation at the data-preprocessing level on predictive accuracy, and examine whether correcting dataset bias leads to more reliable PTB prediction. By comparing results obtained from biased and revised datasets, the study further aims to highlight the critical role of data quality and fairness in building trustworthy AI applications for healthcare.
Methodology
Dataset description
For the development of the clinical decision support system for PTB, a medical dataset of 973 pregnant women was used. The data originated from ‘name of hospital’ as part of the European project ‘name of project’. No missing values existed in the dataset, thus all cases were eligible for the analysis. From a methodological point of view this study had two main objectives: (i) to assess whether data-level bias and representativeness issues exist in the dataset, and (ii) to evaluate how correcting an identified outcome prevalence imbalance affects PTB prediction performance. We begin by examining the dataset to identify any potential data-level bias or representativeness issues in the variables. We then process the dataset to address the detected outcome prevalence imbalance and apply several machine learning algorithms to predict the likelihood of PTB as a binary outcome (i.e., yes/no). We compare the prediction results between the original (biased) and the adjusted (de-biased) datasets, finding statistically significant improvements in the latter using the Wilcoxon signed rank test. We included demographic variables (e.g., maternal age, BMI, height), as well as other clinical variables some of which are known to be associated with PTB (e.g., cervical length, history of PTB, preeclampsia). The gestational age of the women in the dataset was ≥28 weeks (onset of third trimester of pregnancy). The distribution of the gestational age across the dataset is depicted in Figure 1. The dataset consists of both continuous (see Table 1) and binary variables (see Table 2). Distribution of gestational age in the sample. Continuous variables: definitions and descriptive statistics. Binary variables: definitions and percentages of yes and no.
Debiasing
A data-level bias assessment was performed using AI Fairness 360 (AIF360). 11 AIF360 is an open-source library developed by IBM that provides tools for analyzing and mitigating bias in machine learning models and can be applied to datasets with binary and continuous variables. In this study, AIF360 was used as an exploratory tool for dataset-level disparity assessment, rather than for algorithmic or model-level fairness mitigation. The analysis focused on the target variable Preterm_birth and examined various selected attributes (e.g., maternal age, BMI, pregnancy history, etc.) to identify potential disparities in outcome distribution across groups. . The main fairness metrics used are:
Disparate Impact (DI) (Equation (1))
12
: Measures the ratio of favorable outcome rates between unprivileged and privileged groups. A value 0.8 < DI < 1.2 indicates fairness.
Statistical Parity Difference (SPD) (Equation (2))
13
: The difference in the probability of favorable outcomes between the unprivileged and privileged groups. A value close to 0 indicates absence of bias.
DI and SPD values for the variables checked for bias.
During the previous analysis it was observed that the dataset showed a PTB rate of 33%, which is much higher than the usual average rate of 8-10%. 1 Thus, a substantial outcome prevalence imbalance relative to the general population was identified. 15 To reduce the observed PTB rate and bring it closer to the real-world rate, random removal of samples from the dataset was performed. More specifically, random undersampling of PTB cases was applied as a data pre-processing step. According to statistical standards, to ensure accuracy in the tests and the corresponding conclusions, this random sample removal needed to be applied 30 times. 16 The undersampling procedure was repeated 30 times to improve the robustness of the subsequent comparisons and reduce the dependence of the results on a single random draw. The revised dataset that was created after the random removal of PTB cases has a PTB rate of 8.73% and contains 800 records.
Subsequently, a new bias assessment was performed on the socioeconomic variables (i.e., maternal age, smoking status) of the new datasets 17 using AIF360 and the DI and SPD metrics again. Regarding the variable Maternal age, the values of the DI and SPD metrics across all datasets fall within ranges that indicate the absence of bias. In contrast, for the Smoking variable, only 8 out of the 30 generated datasets were found to be unbiased based on the Disparate Impact (DI) metric. Among the remaining 22 datasets, 13 had borderline DI values ranging from 0.70 to 0.76 — very close to the established threshold of 0.8 — while the other 9 had clearly biased values between 0.58 and 0.67. As these datasets were generated after the removal of a significant number of samples, a further correlation analysis was conducted between Smoking and PTB using the Chi-square and Fisher’s Exact Test methods. Statistically significant association was observed in 7 out of the 30 datasets (23%), all of which belonged to the group with lower DI values (0.58–0.67). Additionally, across all datasets, the Statistical Parity Difference (SPD) metric ranged from −0.12 to 0.01, remaining very close to the neutrality threshold of 0. This further supports the conclusion that the Smoking variable does not substantially reinforce disparities against specific groups in the outcome and can therefore be reasonably considered unbiased. Therefore, the adjustment applied in this study should be interpreted as data-level correction of outcome prevalence imbalance, rather than model-level fairness mitigation. The post-adjustment evaluation focused on predictive performance, not on fairness in model predictions across demographic groups.
Preterm birth prediction
Before predicting PTB, data pre-processing was performed to ensure the reliability of the results. Due to modifications in the dataset during the debiasing procedure, the PTB variable contained incorrect NaN values, which were corrected by converting it to integer type and removing erroneous records. Furthermore, the GA (gestational age at delivery) variable was removed because it directly revealed the value of the dependent variable (PTB) and introduced bias into the prediction (i.e., GA≤37 means preterm birth). No imputation was required, as the dataset contained no missing values after preprocessing.
In addition, continuous variables were standardized using StandardScaler prior to training non-tree-based models (e.g., Logistic Regression, Support Vector Machine, and K-Nearest Neighbors), ensuring that feature scale did not influence model performance. Categorical variables were encoded as binary indicators.
For training and internal validation, the dataset was split into two subsets in a 70%-30% ratio (training and test set, respectively). This split was performed before applying the Synthetic Minority Over-samplingTechnique (SMOTE) balancing method to prevent artificial information from being introduced into the test set. All models were trained on the balanced training set derived after the debiasing procedure, while evaluation was performed on a separate hold-out test set. Finally, a stratified 10-fold cross-validation was applied to prevent overfitting while preserving class distribution across folds, with slight adjustments depending on the algorithm used. Random seeds were fixed where applicable to ensure reproducibility.
In order to calculate accurate predictions of PTB, several machine learning models were tested. When selecting these algorithms, the aim was to cover a broad range of modeling approaches in order to achieve comprehensive evaluation of predictive performance, robustness, and generalization in the clinical dataset. Linear models, such as Logistic Regression, are transparent and can be interpreted easily, which is very important in medical applications. Non-linear models, such as Support Vector Machines and Decision Trees, are efficient in identifying complex relationships in the data. Ensemble methods such as Random Forest, Gradient Boosting, AdaBoost, XGBoost, and Bagging were used due to their ability to improve accuracy and reduce overfitting by combining multiple models. Naive Bayes is computationally efficient, especially with categorical features and smaller datasets. Finally, ensemble strategies such as Voting and Stacking leverage the strengths of individual models to further enhance prediction performance and reliability.
Specifically, More specifically: i) Random Forest, 18 an ensemble method that builds multiple decision trees and combines their outputs to improve accuracy and control overfitting. It can handle large datasets of high dimensionality and offers a good balance between bias and variance. ii) Support Vector Machine (SVM), 19 a non-linear model that uses the kernel trick to handle complex data distributions. It is efficient with high-dimensional data and requires minimal memory, making it suitable for complex and diverse obstetric datasets. iii) Logistic Regression, 20 a linear model for binary classification that provides probabilistic predictions. It was chosen for its simplicity and inherent interpretability. iv) K-Nearest Neighbors (KNN), 21 a simple instance-based algorithm that classifies a data point based on the majority class among its closest neighbors. It is easy to implement and good as a baseline for comparing more complex models. v) Decision Tree, 22 which uses a tree-like structure of decisions and consequences to classify the data. It is straightforward to understand and interpret, making it suitable for initial data exploration. vi) Gradient Boosting, 23 an ensemble technique that builds models sequentially, with each new model correcting the errors of the previous. It is a powerful tool for improving accuracy. vii) AdaBoost (Adaptive Boosting), 24 an ensemble method that combines multiple weak classifiers into a strong one by focusing on the most difficult instances to classify. viii) XGBoost (Extreme Gradient Boosting), 25 an advanced gradient boosting algorithm known for its efficiency and high performance on structured data. It handles missing data well and offers extensive hyperparameter tuning. ix) Naive Bayes, 26 a probabilistic classifier based on the Bayes theorem, assuming independence among the predictive variables. It is very efficient in small datasets with categorical features. x) Bagging 27 (Bootstrap Aggregating), 28 an ensemble method that improves stability and accuracy by training mul-tiple models on different subsets of the data. Reduce variance and prevent overfitting. xi) Model Ensemble Methods, 29 such as Voting and Stacking, 30 were tested for further performance improvement. Voting combines predictions from multiple models, leveraging the strengths of each. Stacking uses the predictions of various models as input to a higher-level model that makes the final prediction.
Hyperparameter settings were selected based on established practices and prior literature, with targeted adjustments to improve convergence and model stability. Specifically, Logistic Regression was implemented with max_iter = 1000 and class_weight = “balanced”, Random Forest with 100 trees and class_weight = “balanced”, and K-Nearest Neighbors with n_neighbors = 5. Other models were applied using default configurations with controlled randomization where applicable.
Regarding feature relationships, potential associations between selected variables and PTB were examined using Chi-square and Fisher’s Exact tests, ensuring that observed relationships were statistically and clinically meaningful.
To evaluate the performance of the classification models tested in the present study, six fundamental evaluation metrics were used. These include ROC-AUC, PR-AUC, Accuracy, Precision, Recall, and the F1 score. The ROC-AUC metric reflects the optimal balance between Sensitivity and Specificity, while PR-AUC captures the trade-off between Precision and Recall. These metrics were computed during cross-validation on the balanced training set and subsequently validated on the independent test set.
Results
The performance of the algorithms applied for PTB prediction can be summarized as follows: The best performing ensemble models, Gradient Boosting, AdaBoost, XGBoost, and the Optimized Voting Classifier, consistently achieved the best results, combining high accuracy with balanced metrics. These models demonstrated strong recall (up to 0.98) and F1- scores reaching 0.91, making them well-suited for tasks requiring reliable identification of positive cases. Random Forest also showed excellent performance across all metrics, with particularly high recall (0.96) and accuracy (0.96), reinforcing its suitability for this prediction task. Among simpler models, Logistic Regression achieved notably high recall (0.98) and F1-score (0.91), while K-Nearest Neighbors (KNN) offered good accuracy (0.93) and recall (0.83) but exhibited lower precision (0.66), indicating some false positive predictions. Decision Tree delivered consistent but moderate results for precision, recall, and F1 (all around 0.85), although it was outperformed by ensemble techniques. Naive Bayes, despite its speed and ease of implementation, underperformed in critical metrics (recall: 0.60, F1-score: 0.67) and is not recommended for this application. Finally, the Bagging Classifier benefited from aggregating simple learners, achieving performance comparable to the top models with an F1-score of 0.89 and high precision and recall. In contrast, the Support Vector Machine (SVM), while achieving perfect precision (1.00), failed to detect most positive cases (recall: 0.11), resulting in a very low F1-score and making it unsuitable for this analysis. The results are summarized in Figures 2 and 3 and Table 4. ROC curve comparison among all tested classifiers. Precision-recall curve comparison among all tested classifiers. Prediction results for PTB.
Prediction results for PTB in the original dataset.
To investigate whether the observed improvement in the results when using the debiased dataset is statistically significant, the Wilcoxon signed rank test was applied. The choice of this nonparametric test was based on the relevant methodological justification by Demsˇar (2006), 31 who recommends its use for statistical comparison of classifiers across different datasets when parametric assumptions such as normality of the differences are not met. In the current analysis, the four key metrics (Accuracy, Precision, Recall, and F1-score) were compared for each algorithm separately between the two datasets. That is, for each metric, the performance of the same algorithms was assessed statistically significantly when applied to the original and the new datasets, respectively.
The test results showed statistically significant differences (p-value < 0.05) for Accuracy (p-value = 0.0254), Recall (p-value = 0.0029), and F1-score (p-value = 0.0020), supporting the hypothesis that classifier performance improved in the new dataset. In contrast, for the Precision metric (p-value= 0.0742), the difference was not considered statistically significant, indicating that the accuracy of positive predictions did not change substantially between the two data sets. Nevertheless, the p-value for Precision is close to the conventional threshold of 0.05, which can be viewed as a borderline indication of a possible difference, though not strong enough to be confirmed with the available data. 32
These conclusions can be explained by the change in the dataset, where the original dataset was reduced and updated. Specifically, the reduction involved the random removal of data causing sampling bias, which appears to have allowed the algorithms to learn more representative and reliable features. 33 The simultaneous improvement in both Recall and F1- score indicates that the overall ability of the classifiers to correctly identify the positive class has increased. In imbalanced problems like the present one, it is common for an increase in Recall to be accompanied by a slight decrease in Precision, as additional positive predictions may include some false positives. In this context, 34 the F1-score metric is preferred as a more representative measure of true performance, since it balances both Precision and Recall. 35 The borderline non-significant difference in Precision suggests that although the new dataset improved overall performance, the models’ ability to avoid false positive predictions remained stable.
Although the best Precision value was slightly lower in the revised dataset than in the original one, this likely reflects the well-known trade-off between Precision and Recall. After correcting the outcome prevalence imbalance, the models became more sensitive to PTB cases, increasing Recall while also producing a slightly higher number of false positive predictions. In the context of PTB prediction, this trade-off may be clinically acceptable, since missing a true PTB case can be more critical than generating additional false positives.
Note that Net Reclassification Improvement (NRI) and the Integrated Discrimination Improvement (IDI) could not be applied as these metrics require both models to be applied to the same validation set, yielding paired, subject-level predicted probabilities. In our study, the models were trained and evaluated on different analytic versions of the dataset (unbiased/balanced), which differ in sample size and composition as the balanced dataset contained fewer positive cases. As a result, predictions are not available for an identical set of individuals, and reclassification indices cannot be calculated validly. 36
The general picture of the results, with significant improvements in most critical metrics and positive changes for most classifiers, confirms the effectiveness of the debiasing procedure and the creation of the new data set. However, it is important to note that not all algorithms exhibited performance improvement across all metrics. For instance, Support Vector Machine (SVM) maintained perfect precision but continued to show extremely low recall, resulting in a poor F1-score. Similarly, for some models, such as Naive Bayes, performance remained suboptimal despite the dataset refinement. These cases indicate that certain classifiers may not be well-suited to the specific structure and imbalance of the dataset, even after debiasing. Thus, while the overall trend points toward improvement, individual algorithm behavior must be considered in model selection.
Discussion
This paper, aims to predict the occurrence of PTB while considering data-level bias and representativeness issues in the original dataset. In this context several machine learning algorithms were employed for both binary classification (PTB vs. non-PTB). The prediction process utilized an existing dataset that included demographic information and a wide range of laboratory and clinical parameters. In this work, the original dataset was preprocessed to address an outcome prevalence imbalance, resulting in the creation of a new dataset of smaller size but equivalent informational content.
Related work
There are several studies in the literature that apply AI and ML to predict PTB. The study by Vovsha et al. 37 focused on uniparous women, applying logistic regression and SVM, with SVM outperforming in predicting high-risk cases. In the work by Begum et al., 38 3,500 pregnant women were examined using K- NN, SVM, and decision trees, with the Decision Tree showing the highest accuracy (99%). At the same time, Koivu and Sairanen 39 utilized approximately 16 million records from the US CDC, including demographic and pregnancy-related data. The preprocessing excluded minors, multiple pregnancies, infections, gestational age below 21 weeks, and the parity variable. They tested logistic regression (optimized with BFGS), LightGBM for fast and accurate handling of large datasets, and two neural network architectures: a two-layer deep neural network with Leaky ReLU and a self-normalizing neural network (SELU) with four hidden layers. The SELU network performed best. Additionally, they applied ensemble learning techniques to achieve stable and reliable predictions. They also employed ensemble learning methods that maintained consistent performance. AlSaad et al. 40 developed PredictPTB based on BiLSTM neural networks using EHR data, achieving superior accuracy, with significant risk factors including preeclampsia, previous PTB, and cesarean section. The study by Zhang et al., 41 at a hospital in China examined 21 risk factors, with the most important being premature rupture of membranes, multiple pregnancy, and placenta previa. They applied AdaBoost with high performance (F1-score 0.969, AUC 0.93) and external validation showing 100% accuracy for term births and 72.73% for PTBs.
Some studies employed specialized approaches, such as Watson et al. 42 using QUiPP, which predicts PTB within 7 days based on fetal fibronectin and clinical data, reducing unnecessary interventions. Meanwhile, Moreira et al. 43 used Support Vector Machines (SVM) with a linear kernel to predict PTB in hypertensive pregnant women, achieving good accuracy (0.821) and a satisfactory ROC area (0.785). This method is recommended for use in decision support systems and mobile health applications. Additionally, Grzymala-Busse et al., 44 applied the LERS method, based on rough set theory. This approach demonstrated accuracy between 68% and 90%, outperforming traditional methods in predicting PTB. Moreover, a retrospective cohort study 35 using Iran’s IMaN Net birth records (2020–2022; n=8,853) evaluated multiple supervised machine-learning models to predict spontaneous preterm birth (<37 weeks) from socio-demographic, medical/obstetric history, and current pregnancy factors. Although several clinical and demographic variables were associated with PTB, all tested models showed poor discriminatory performance (AUC-ROC 0.57–0.65; best deep learning feed-forward AUC=0.65), indicating that these approaches and available predictors were insufficient for reliable PTB risk prediction and motivating more robust future modeling. However, no bias analysis was conducted.
Finally, studies based on electrohysterographic (EHG) recordings, such as Despotovic’ et al. 45 which used data from the 22nd to 25th week of gestation with spectral and nonlinear feature analysis, showed that Random Forest with ADASYN achieved a sensitivity of 98.4% and an AUC of 99%. Similarly, the study by Narmadha et al., 46 involving preprocessing of EHG signals and algorithms like KNN, LR, and AdaBoost, found that SVM with a polynomial kernel achieved 99.7% accuracy.
Extending the state-of-the-art
This work extends the work by [reference removed for double blind review], where data from 973 pregnant women were used to predict PTB. They applied algorithms such as Random Forest, XGBoost, and Optimized Voting, with the latter achieving an accuracy of 94% and recall of 97%. They also utilized SHAP for model interpretation. In contrast to previous studies that mainly focus on predictive performance, the present study additionally examined how dataset representativeness and target outcome distribution affect prediction results. This study approaches the prediction of PTB through one main innovation, as before applying machine learning algorithms, a systematic data-level bias assessment of the available dataset was conducted, something not mentioned in previous studies. More specifically, the study examined whether the original dataset exhibited distortions in the distribution of the target outcome and whether correcting this issue would affect model performance.
Therefore, the various algorithms were executed on an adjusted dataset with close to real-world PTB prevalence, which led to statistically significant improvements in most of the predictive metrics and supports the importance of careful dataset curation before model development. Through this different approach, the study sought to increase the reliability of the predictions and the clinical applicability of the results. These findings suggest that, in clinical decision support settings, attention should not be limited to algorithm selection alone, but should also include the quality, representativeness, and preprocessing of the underlying data.
Overall, this work highlights the importance of rigorous examination and preprocessing of training data to improve model performance. This is particularly important in healthcare applications, where the credibility and practical adoption of AI systems depend highly on the robustness, interpretability, and clinical plausibility of their results.
Limitations
A number of limitations related to this study need to be acknowledged. First, the proposed de-biasing methodology addressed the existence of bias at the data-preprocessing level. Specifically, it corrected the outcome prevalence imbalance, and did not include algorithm-level or post hoc fairness mitigation techniques. Therefore, fairness in model predictions across demographic or protected groups was not directly evaluated using standard criteria such as equalized odds or demographic parity.
Second, the revised dataset was created through random undersampling of PTB cases. This approach improved the representativeness of the target outcome distribution. However, it also reduced the overall sample size which may have resulted in the loss of potentially useful information.
Third, the dataset was collected from one clinical setting. As a result, the generalizability of the findings to other hospitals, regions, or patient populations remains uncertain. In order for broader clinical conclusions to be drawn, an external validation on a more diverse cohorts is necessary.
Conclusions and future work
This study highlighted the potential of applying machine learning techniques for the prediction of delivery time focusing on the prediction of PTB giving special attention in the dataset quality and the correction of bias. The identification and correction of outcome prevalence imbalance in the original dataset, along with the creation of a balanced dataset, significantly contributed to the improvement of the results, compared to the initial biased dataset. These results highlight the importance of careful data pre-processing. In conclusion, this work demonstrated that machine learning can serve as a valuable tool in the field of prenatal care, provided it is supported by sufficient and high-quality data and class balance. Overall, the findings also highlight the importance of dataset quality and representativeness for developing reliable and clinically applicable AI-based tools. In future work, we aim to include the collection and integration of more and higher quality features (e.g., medical examinations, family history, pregnancy progress data), as well as the use of interpretability tools (such as SHAP) to enhance transparency in the results. Moreover, we aim to expand our dataset by collecting data from different sites. Finally, we aim to evaluate fairness in model predictions across demographic groups.
Footnotes
Ethical considerations
This study was submitted and approved by the Research Ethics and Conduct Committee of the Aristotle University of Thessaloniki (94521/2022).
Consent to participate
All patients that participated in this study provided written informed consent.
Author contribution
Methodology, analysis and writing, E.M., E.R. A.B.; review and validation, A.B.; data collection and curation and clinical problem definition, I.T. and T.D.; supervision N.B and P.D.B.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article:This work was funded by the European Union’s Horizon 2020 Innovation Action programme, ‘HosmartAI - Hospital Smart Development Based on AI,’ under Grant Agreement No. 101016834.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
The dataset generated during the current study will be available from the corresponding author on reasonable request.