Risk of bias in machine learning and statistical models to predict height or weight: a systematic review in fetal and...

Area of Science:

• Pediatrics
• Biostatistics
• Medical Informatics

Background:

• Accurate prediction of fetal and child growth is crucial for early intervention and improved outcomes.
• Existing prediction models, some over 30 years old, are still recommended by national bodies, necessitating an evaluation of their current validity and potential biases.
• This systematic review addresses the risk of bias in statistical and machine learning models used for predicting height or weight in individuals under 20 years of age.

Purpose of the Study:

• To investigate the risk of bias in statistical and machine learning models for predicting fetal, infant, and child height or weight.
• To inform the current standard of research in growth prediction.
• To provide insights into the continued use of outdated prediction equations.

Main Methods:

• Systematic search of MEDLINE and EMBASE for peer-reviewed original research studies published in 2022.
• Inclusion of studies developing or validating multivariable models (≥2 variables) for height/weight prediction, excluding imaging, genetics, or metabolomics.
• Risk of bias assessment for all models using the Prediction model Risk Of Bias ASsessment Tool (PROBAST).

Main Results:

• Sixty-four studies were included, assessing 180 developed and 61 validated models.
• Sample size was inadequately considered in model development (10%) and validation (13%).
• A significant majority (77%) of developed models predicted a dichotomized outcome, despite height and weight being continuous variables.

Conclusions:

• The review identified significant risks of bias in current models for predicting fetal and child growth.
• Inadequate consideration of sample size and inappropriate dichotomization of continuous outcomes are prevalent issues.
• These findings highlight the need for improved methodologies and updated standards in growth prediction research.