Appraisal of Gene Expression-Based Classifiers for Neuropsychiatric Disorders: A Meta-Regression

A substantial body of research examines the potential of gene-expression-based biomarkers for diagnosing and selecting treatments for neuropsychiatric disorders, yet no clear consensus has been reached regarding the influence of controllable factors such as study design and model selection on the performance of gene-expression-based classifiers. To investigate study characteristics and methodologies that influence the accuracy of studies using transcriptomics to classify neuropsychiatric disorders, we conducted a literature review and meta-regression of relevant studies. We extracted several characteristics from each study, including the number of samples in a training dataset, approach for model validation, and classification model. Using univariate and multivariate mixed-effect meta-regression analyses, we estimated the association between these study characteristics and reported classification accuracies. Machine Learning (ML) models accounted for 55% of all models; Deep Learning (DL) models accounted for 20%, and variations of Logistic Regression models made up the remaining 25%. Support vector machine (SVM) was the most common model type (17%). The use of withheld test samples (56%) was the most frequent approach for validating the performance of classification models. We found significant associations between reported accuracies and study-rated bias risk, model type, class ratio, and validation approach. Overall, this review provides helpful insights into study characteristics that significantly influence classification accuracies and emphasizes the importance of prudent methodologies for training and evaluating classification models to mitigate biased accuracy estimates. In future research, these findings can direct study designs and support the development of clinically reliable, minimally invasive gene-expression biomarkers to improve diagnostic workflows and patient outcomes in neuropsychiatry.