Use of uterine activity to predict preterm birth by artificial intelligence assisted models: a narrative systematic...

The benefit of interventions to improve neonatal outcomes of preterm birth (PTB) must be balanced with the associated fetal and maternal risks. Artificial intelligence (AI) could be used to assess uterine contractions and consequently help to predict PTB. This paper aims to assess the predictive accuracy and applicability of AI models currently using uterine contractions in PTB prediction.

A systematic Embase, Medline, Pubmed and Web of Science review was conducted using PRISMA guidelines. Eligible studies assessed EHG or time-series data using AI methods, including deep learning/machine learning/neural networks to predict PTB. Data on AI model performance measures, validity, and applicability were collected. Results are reported as a narrative review due to study heterogeneity. Bias was assessed using the PROBAST framework.

The studies used various Electrohysterography (EHG) contractility features and/or classifiers for AI analysis and varying performance measures to assess predictive accuracy for PTB. A wide range of EHG features were assessed included temporal, spectral, entropy and topological features. A total of 53 records were identified for inclusion. Of these, 18 examined EHG features, 22 assessed AI classifiers, and 3 tested both. Excellent classification performance (ACC and/or AUC ≥0.9) were reported by 38.8% (7/18) of studies examining EHG features and 86.3% (19/22) of studies assessing AI classifiers. Non-linear features outperformed linear features, and deep-learning models such as neural networks were the highest-performing classifiers. Bias assessment showed 86.7% (46/53) had an unclear or high risk of bias. Key concerns include unbalanced data, small sample size and lack of validity outside of sampled datasets.

Non-linear features and DL models offer superior results. However, we did not find evidence of external validation, thus the applicability of models in uterine contraction assessment for the prediction of PTB remains limited. Future research requires an emphasis on clinical data integration within high-quality studies, as well as more more studies focusing on early PTB detection.