Genomic Machine Learning Meta-regression: Insights on Associations of Study Features with Reported Model Performance

BackgroundMany studies have been conducted with the goal of correctly predicting diagnostic status of a disorder using the combination of genetic data and machine learning. The methods of these studies often differ drastically. It is often hard to judge which components of a study led to better results and whether better reported results represent a true improvement or an uncorrected bias inflating performance. MethodsIn this systematic review, we extracted information about the methods used and other differentiating features in genomic machine learning models. We used the extracted features in mixed-effects linear regression models predicting model performance. We tested for univariate and multivariate associations as well as interactions between features. ResultsIn univariate models the number of hyperparameter optimizations reported and data leakage due to feature selection were significantly associated with an increase in reported model performance. In our multivariate model, the number of hyperparameter optimizations, data leakage due to feature selection, and training size were significantly associated with an increase in reported model performance. The interaction between number of hyperparameter optimizations and training size as well as the interaction between data leakage due to optimization and training size were significantly associated reported model performance. ConclusionsOur results suggest that methods susceptible to data leakage are prevalent among genomic machine learning research, which may result in inflated reported performance. The interactions of these features with training size suggest that if data leakage susceptible methods continue to be used, modelling efforts using larger data sets may result in unexpectedly lower results compared to smaller data sets. Best practice guidelines that promote the avoidance and recognition of data leakage may help the field advance and avoid biased results.