Can combining cervical imaging, uterine electrical activity, and biomarkers with artificial intelligence improve the prediction of preterm birth? Promises, performance figures, and the chronic problem that shadows this field: data imbalance and poor external validation.
⚕️ Target audience: This content is aimed at healthcare professionals and is for educational/current-information purposes. AI tools offer clinical decision support; It does not replace clinician responsibility and approved clinical workflows.
Preterm birth (birth occurring before 37 weeks of gestation) is the leading cause of neonatal mortality and morbidity worldwide; It affects approximately 13.4 million births annually. Despite decades of research, overall preterm birth rates remain largely unchanged, and classical prediction methods—risk factor scoring, cervical length measurement, and fetal fibronectin testing—show limited performance in differentiating individual risk (Henricks & Nelson, Semin Perinatol 2026). Artificial intelligence (AI), especially machine learning (algorithms that learn patterns from data) and deep learning (multilayer artificial neural networks), have rapidly entered the literature in recent years with the claim of filling this gap.
This paper examines AI-based preterm prediction across four primary data sources: cervical imaging, uterine electrical activity (electrohysterography), biomarkers (quantitative fetal fibronectin), and multivariable clinical models. It also honestly exposes the problem of dataset imbalance and poor external validation—the most critical weaknesses of this field.
Cervical length on transvaginal ultrasonography is the single most established marker of spontaneous preterm birth in asymptomatic and symptomatic pregnant women; however, its predictive value is modest. Deep learning has the potential to extract additional texture information from these images that the human eye cannot capture.
The strongest evidence in this area comes from a multicenter study in Denmark with data collected at 19 centers (Sejer et al., Am J Obstet Gynecol 2026). Trained on 4224 pregnancies and 7862 cervical ultrasound images, the deep learning model outperformed cervical length in predicting spontaneous preterm birth <37 weeks: sensitivity 0.51 versus 0.41 (p<0.001), area under the curve (AUC — measure of discriminatory power) 0.75 versus 0.67 (p<0.001) when specificity was fixed at 0.85. In late preterm births (34–37 weeks) the model showed a 36.6% higher sensitivity. Remarkably, saliency maps (heat maps showing which pixels the model based its decision on) revealed that in 70% of the cases, the model focused on the inner line of the lower uterine segment — meaning the model “sees” more than just the cervical length. This is one of the rare studies with large, real-world data; however, an AUC level of 0.75 is still a modest discriminatory power for individual clinical decisions.
Automated measurement of cervical length is also an active area of research; Deep learning algorithms aim to segment the cervix in transvaginal images and standardize measurement, thus reducing operator variability (Kwon et al., IEEE J Biomed Health Inform 2024). Additionally, elastography (ultrasound technique that measures tissue stiffness) and quantitative tissue-structure (radiomics) features are being tested as input to AI models; however, most of these approaches are small, single-center, and lack prospective evidence.
Electrohysterography (EHG—uterine electrical activity recorded from the maternal abdominal surface, i.e., uterine EMG) holds the promise of non-invasively capturing the myometrial activity underlying preterm labor. While tocodynamometry gives only the mechanical reflection of contractions, EHG directly measures the electrical source. Signal processing (wavelet transform, empirical mode decomposition) and machine learning (support vector machines, neural networks) have been extensively applied to extract preterm/term distinguishing features from these signals (Barnova et al., Artif Intell Med 2026).
The overwhelming majority of research in this area relies on a single publicly available data set: the Term-Preterm EHG Database (TPEHG DB). This set, collected at the University Medical Center Ljubljana between 1997–2005, contains 300 EHG records recorded around the 22nd or 32nd week of gestation—262 of which resulted in term and only 38 in preterm birth (PhysioNet TPEHG DB). That is, only ~13% of the data set belongs to the preterm class. This imbalance is the source of the methodological pitfalls discussed below.
An honest comparison: In the study where the Amsterdam group applied end-to-end deep learning (LSTM and temporal convolutional networks) on TPEHG data, the resulting AUC was only 0.58 — roughly on par with models using hand-crafted features. Moreover, adding existing clinical data to the EHG did not improve performance. The authors clearly emphasized that a dataset of high-risk patients should be collected to reduce false positives (Fischer et al., Comput Biol Med 2023).
There are also studies in the literature reporting AUC exceeding 90%; but these figures are mostly the product of methodological artifact. We explain why in the next section.
The inherent obstacle to preterm prediction models is the inherently unbalanced preterm/term ratio. Most studies have tried to solve this problem with the synthetic minority over-sampling technique (SMOTE — a method that generates artificial samples for the minority class). The problem is that if synthetic data is generated before the dataset is split into training/testing, information leakage occurs between the training and testing partitions; The classification task becomes artificially easier and the true generalization capacity of the model appears overestimated.
The study of the Valencia group that directly examined this pitfall is instructive (Nieto-Del-Amor et al., Sensors 2022): when the sampling strategy was set up correctly — when artificial data was applied only at the appropriate stage and post-splitting — an AUC level of 94.5 ± 4.6% on TPEHG could be achieved. However, the authors made clear that many high-performance reports in the literature overestimate true generalization ability precisely because of this leakage. So the gap between “correct” and “wrong” methodology on the same data set is evidence of how fragile the reported performance is.
Conclusion of systematic evidence: In a recent systematic review of 53 records evaluating uterine activity-based AI models, 86.7% of studies were at high or unclear risk of bias (according to the PROBAST tool). Key problems: unbalanced data, small sample, and failure to demonstrate validity outside the sampled data set. The review shows that non-linear features and deep learning models provide superior results; however, they concluded that no external validation evidence was found and therefore clinical applicability remained limited (Chopra et al., J Matern Fetal Neonatal Med 2026).
Multivariate models that combine biomarkers with clinical risk factors, rather than images or signals alone, are the closest approach to clinical reality today. The most mature example of this is the QUiPP approach based on quantitative fetal fibronectin (qfFN — numerical measurement of fibronectin density in cervicovaginal fluid).
In the QUIDS study, a risk model including qfFN along with clinical variables such as smoking, ethnicity, nulliparity, and multiple pregnancy was developed and externally validated. In predicting spontaneous preterm birth within 7 days of testing in symptomatic women, the model demonstrated AUC 0.89 in the development cohort and AUC 0.89 in the independent UK cohort (Stock et al., PLoS Med 2021). One important caveat: calibration was excellent in the low-risk range, but risk was underestimated in high-risk women—recalibration of the model intercept was required.
The same limitation was repeated in a more recent external validation (Fischer et al., Ultrasound Obstet Gynecol 2025): in three independent European cohorts, the QUiPP App v.2 performed with an AUC of 0.74–0.84 for preterm birth within 1 week with the combination of cervical length + qfFN, with a peak AUC of 0.91 for prediction of <30 weeks. Here the pattern was the same: while the calibration was reliable in low predicted risk patients, the risk was systematically underestimated compared to the true event rate in high risk women. This also shows that models are more reliable for short-term (e.g. <30 weeks) results than for long-term (<37 weeks) results. Although broader “multi-omics” (integration of genomic, proteomic, metabolomic data) and transformer architectures are promising, they are still in the early research stages (Henricks & Nelson, Semin Perinatol 2026).
Reported AUC values across modalities range from approximately 0.61 to 0.89 (Lee, Obstet Gynecol Sci 2026). But the methodological quality behind these figures is poor, and this is the field's most honest summary:
On the other hand, positive signals are also real: a large, multicenter cervical imaging study covering different ultrasound devices (Sejer 2026) and the externally validated qfFN-based QUiPP model (Stock 2021; Fischer 2025) show that AI can surpass classical single-markers when methodologically correct. The problem is not in the technology itself, but in the methodological rigor of the studies in the field.
Today, the AI-assisted tools closest to clinical use in predicting preterm birth are not EHG classifiers but externally validated multivariable biomarker models (qfFN + cervical length + clinical factors; e.g. QUiPP). Even these should be used where they are most powerful—in safely excluding low risk and in short-term forecasting; Positive results should be interpreted with caution due to a tendency to underestimate high risk. EHG-based deep learning models, although exciting, are still in the research phase without external validation and prospective evidence.
Prior to transition to clinical practice, the priority needs are clear (Lee 2026; Devoe 2025): compliance with TRIPOD-AI reporting standards, multicenter and external validation, reporting of calibration metrics, and ultimately prospective demonstration of clinical benefit. Clinicians should ask three questions when evaluating an AI tool: What population was it trained on? Has it been externally validated in an independent cohort? Has its calibration—not just its discriminatory power—been reported? A model that does not meet these three criteria cannot be used safely at the bedside, even if it reports an impressive AUC.
Last updated: . Sources have been confirmed with their publication dates and current versions.