The importance of discriminative power rather than significance when evaluating potential clinical biomarkers in epilepsy research.

Abstract

The quest for epilepsy biomarkers is on the rise. Variables with statistically significant group-level differences are often misinterpreted as biomarkers with sufficient discriminative power. This study aimed to demonstrate the relationship between significant group-level differences and a variable's power to discriminate between individuals. We simulated normal-distributed datasets from hypothetical populations with varying sample sizes (25-800), effect sizes (Cohen's d: .25-2.50), and variability (standard deviation: 10-35) to assess the impact of these parameters on significance and discriminative power. The simulation data were illustrated by assessing the discriminative power of a potential real-case biomarker-the EEG beta band power-to diagnose generalized epilepsy, using data from 66 children with generalized epilepsy and 385 controls. Additionally, we evaluated recently reported epilepsy biomarkers by comparing their effect sizes to our simulation-derived effect size criterion. Group size affects significance but not discriminative power. Discriminative power is much more related to variability and effect size. Our real data example supported these simulation results by demonstrating that group-level significance does not translate, one to one, into discriminative power. Although we found a significant difference in the beta band power between children with and without epilepsy, the discriminative power was poor due to a small effect size. A Cohen's d of at least 1.25 is required to reach good discriminative power in univariable prediction modeling. Slightly over 60% of the biomarkers in our literature search met this criterion. Rather than statistical significance of group-level differences, effect size should be used as an indicator of a variable's biomarker potential. The minimal required effects size for individual biomarkers-a Cohen's d of 1.25-is large. This calls for multivariable approaches, in which combining multiple variables with smaller effect sizes could increase the overall effect size and discriminative power.