Methodological conduct and risk of bias in studies on prenatal birthweight prediction models using machine learning techniques: a systematic review.

Objective: To assess the methodological quality and the risk of bias, of studies that developed prediction models using Machine Learning (ML) techniques to estimate prenatal birthweight.

Study design and methods: We conducted a systematic review, searching the PubMed databases between 01/01/2018 and 01/08/2022, for studies that developed fetal weight prediction models using ML. We used the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement to assess the reporting quality of included publications and the Prediction Model Risk of Bias Assessment Tool (PROBAST) to assess the risk of bias. We measured the overall adherence to the TRIPOD reporting checklist, provided a detailed analysis of the methodological quality of each study, and examined risk of bias in specific domains, including participant, predictor, outcome and analysis.

Results: Fourteen studies were included and the adherence to the TRIPOD reporting items ranged from 34.62% to 80.77%, with a median adherence of 63.19%. The studies showed significant variation in their methodological rigor, with a particularly high risk of bias in the selection of participants and predictors. Notably, issues related to missing data, sample size adequacy, performance evaluation, and model validation were prominent across studies. Several studies showed limited model transparency and reproducibility.

Conclusion: Methodological quality of the ML-based prediction models for prenatal birthweight estimation was generally poor, with most studies at high risk of bias. There is an urgent need for improvements in the design and reporting of these studies. The adaptation of the TRIPOD and PROBAST statements specifically for ML models should be promoted to enhance transparency and reproducibility, which would facilitate the wider clinical application of ML-based prediction models and reduce research waste.