| Popis: |
There is significant interest in using neuroimaging data to predict behavior. The predictive models are often interpreted by the computation of feature importance, which quantifies the predictive relevance of an imaging feature. Tian and Zalesky (2021) suggest that feature importance estimates exhibit low test-retest reliability, pointing to a potential trade-off between prediction accuracy and feature importance reliability. This trade-off is counter-intuitive because both prediction accuracy and test-retest reliability reflect the reliability of brain-behavior relationships across independent samples. Here, we revisit the relationship between prediction accuracy and feature importance reliability in a large well-powered dataset across a wide range of behavioral measures. We demonstrate that, with a sufficient sample size, feature importance (operationalized as Haufe-transformed weights) can achieve fair to excellent test-retest reliability. More specifically, with a sample size of about 2600 participants, Haufe-transformed weights achieve average intra-class correlation coefficients of 0.75, 0.57 and 0.53 for cognitive, personality and mental health measures respectively. Haufe-transformed weights are much more reliable than original regression weights and univariate FC-behavior correlations. Intriguingly, feature importance reliability is strongly positively correlated with prediction accuracy across phenotypes. Within a particular behavioral domain, there was no clear relationship between prediction performance and feature importance reliability across regression algorithms. Finally, we show mathematically that feature importance reliability is necessary, but not sufficient, for low feature importance error. In the case of linear models, lower feature importance error leads to lower prediction error (up to a scaling by the feature covariance matrix). Overall, we find no fundamental trade-off between feature importance reliability and prediction accuracy. |
