AI predicts preterm birth risk with 82% accuracy

Could AI predict preterm births before symptoms arise? A new study finds that machine learning models, especially SVMs, can assess risk with impressive accuracy—offering hope for earlier interventions and better neonatal outcomes.

Study: Predicting preterm birth using machine learning methods. ​​​​​​​Image Credit: StockKK / Shutterstock

In a recent study in the journal Scientific Reports, researchers evaluated the accuracy, precision, and F1-score of several machine learning (ML) models in predicting the likelihood of preterm births in 50 pregnant women. Despite several attempts at unraveling the underlying causes of preterm birth, the multifaceted nature of the condition has made identifying a biological cue for preterm births hitherto impossible.

Given its status as a significant health concern and its strong correlation to adverse neonatal outcomes (mortality and morbidity), this study aims to use ML models to predict preterm risk, thereby allowing for timely interventions in high-risk women. Study findings identified linear support vector machines (SVMs), particularly those with optimized hyperparameters, as the best-performing (accuracy = 82%) out of the several (n = 5) tested.

Background

Preterm births, also called 'premature births,' are when babies are born before 37 weeks of pregnancy. They can be medically severe conditions substantially increasing neonatal complications, including breathing difficulties, feeding difficulties, cerebral palsy, and even neonatal mortality. Unfortunately, preterm births are an increasingly common occurrence, with the World Health Organization (WHO) estimating that 1 in every 10 babies is born premature (WHO 2020).

While decades of research have elucidated some of the underlying causes of preterm birth, including maternal smoking, alcohol consumption, stress, pollution exposure, and, most recently, genetics, the complex interplay between these factors has resulted in a lack of a single, definitive risk determinant of the condition. Consequently, clinicians today rely on risk evaluation models to determine the likelihood of preterm birth and administer timely interventions and care.

Machine learning models (ML) are witnessing unprecedented use in clinical decision support systems, including risk determination. Their ability to detect patterns invisible to traditional statistics and leverage a wide range of input data types (transvaginal ultrasound, electronic health records (EHRs), and electrohysterogram signals) makes them increasingly valuable in preventive medicine. While ML models have been studied before for preterm birth prediction, the present study focuses on identifying the most effective models and improving their predictive accuracy through hyperparameter tuning.

About the Study

The present study aims to identify the best-performing ML algorithms in determining preterm risk by leveraging a cohort of 50 women (28 cases and 22 controls) to assess their accuracy metrics. Participant data was obtained from pregnant women admitted to Dr. Antoni Biziel University Hospital in Bydgoszcz, Poland. Study data included detailed medical examinations (health evaluations, gynecological assessments, and blood tests) and medical questionnaires (participants' medical history, current medication, and other clinically relevant details).

This study evaluated several cutting-edge ML algorithms, including XGBoost, logistic regression, CatBoost, decision trees, and support vector machines (SVMs). To maximize the algorithms' F1 scores (and thereby performance), models were subjected to hyperparameter optimization using the Optuna framework. The study then assessed model-specific performance across four main metrics: accuracy, recall, precision, and F1 Score.

To establish statistical significance and differentiate performance between models, chi-squared tests and Welch’s unpaired t-tests were employed. Finally, the best-performing models were subjected to feature performance analysis to help identify participant traits that contributed most to model accuracy, thereby hinting at clinically relevant symptoms that could be used to predict preterm births in future investigations.

Study Findings

The study identified the linear SVM (with optimized hyperparameters) as the best-performing model, achieving 82% accuracy, 86% recall, 83% precision, and an 84% overall F1 score. The linear SVM was followed closely by the logistic regression model (also with optimized hyperparameters), which achieved comparable performance with an 80% accuracy, 82% recall, 82% precision, and 82% overall F1-score. Notably, both of these models are objectively relatively simple algorithms.

More complex algorithms, such as XGBoost and CatBoost, performed more poorly than expected, potentially due to the small dataset size (n = 50), which limited their ability to generalize effectively. The study suggests that these models may have been too complex for the available dataset, leading to inefficiencies in learning from the given features. Elementary models (e.g., random forests and decision trees) also underperformed, not only due to dataset size limitations but also because of their difficulty in handling the large number of features supplied in the study.

Feature performance analysis revealed that in addition to C-reactive protein (CRP) from blood morphology parameters and parity (the number of previous childbirths), hematocrit (HCT) and platelet count (PLT) were also significant predictors of preterm birth. Notably, education level was also identified as a statistically significant factor, suggesting that socioeconomic factors play a role in preterm birth risk. These findings indicate that factors related to inflammation and blood composition play an important role in preterm risk assessment.

"Collectively, these findings suggest that preterm birth is driven by a multifactorial interplay of physiological, socioeconomic, and behavioral factors. The findings highlight the need for integrated care approaches that address both biological and social determinants in pregnancy."

Conclusions

The present study identified linear SVMs as the ML model with the highest accuracy, precision, recall, and overall F1 score among the five models evaluated. Alongside logistic regression (the second-best performer), this model highlights that optimal algorithmic complexity plays a critical role in preterm birth prediction, as models that were either too simple or too complex tended to underperform.

Despite the study's limited sample size (n = 50 participants), which significantly influenced model performance, the findings are promising. However, the researchers caution that larger-scale studies are necessary to validate the models' generalizability. Future research should focus on collecting larger, more diverse datasets and including earlier-stage pregnancy screening to enhance predictive accuracy.

"The results of this study have the potential to inform the development of interventions aimed at reducing the incidence of preterm birth… prospective studies should be designed to explore the real-world applicability of the identified model in clinical settings, where its predictive power could aid in early risk identification and intervention strategies for preterm birth."