Interpreting models interpreting brain dynamics

Abstract

Brain dynamics are highly complex and yet hold the key to understanding brain function and dysfunction. The dynamics captured by resting-state functional magnetic resonance imaging data are noisy, high-dimensional, and not readily interpretable. The typical approach of reducing this data to low-dimensional features and focusing on the most predictive features comes with strong assumptions and can miss essential aspects of the underlying dynamics. In contrast, introspection of discriminatively trained deep learning models may uncover disorder-relevant elements of the signal at the level of individual time points and spatial locations. Yet, the difficulty of reliable training on high-dimensional low sample size datasets and the unclear relevance of the resulting predictive markers prevent the widespread use of deep learning in functional neuroimaging. In this work, we introduce a deep learning framework to learn from high-dimensional dynamical data while maintaining stable, ecologically valid interpretations. Results successfully demonstrate that the proposed framework enables learning the dynamics of resting-state fMRI directly from small data and capturing compact, stable interpretations of features predictive of function and dysfunction.

Towards Greater Neuroimaging Classification Transparency via the Integration of Explainability Methods and Confidence Estimation Approaches

“…We identified a number of brain network pairs useful to the classification of SZ and HCs. Previous studies have found widespread effects of SZ upon the CBN [2][15][27], SMN [15], and SCN [2] similar to our results. Additionally, some studies have identified differences in the VSN/VSN [25][58], VSN/SCN [25][59], and VSN/SMN [25][60] as important to differentiating SZ from healthy individuals.…”

“…This would indicate that the aberrant effects of SZ upon brain network interactions tend to be temporally localized. This finding is supported by [15], which found similar effects of SZ upon attention values of a long short-term memory network. Additionally, this finding is related to those of other studies that have found effects of SZ upon brain network dynamics [24], [26].…”

“…In our temporal analyses, we first sought to understand whether there were differences in the temporal distribution of importance over time between SZs and HCs and next sought to understand whether there were differences in the temporal distribution of importance between the basic model and model with MCBN and MCD. To this end, we adapted a method presented in [15]. We used Earth mover's distance (EMD), a distance measure comparing two densities, to calculate the distance for each participant between the importance of each dFNC feature over time with the average importance across time.…”

“…Nevertheless, if the methods being developed for automated diagnosis are ever to be implemented in a clinical setting, model transparency must be taken into consideration [11]. While there are multiple dimensions to transparency [12], there is little or no available literature on the subject of integrating two of those dimensions - model confidence estimates [13], [14] and model explainability [2][15] – into the same model for neuroimaging classification. In this study, we compare two existing approaches for estimating model confidence [13], [14].…”

Identifying Neuropsychiatric Disorder Subtypes and Subtype-dependent Variation in Diagnostic Deep Learning Classifier Performance

“…We only examined spatial differences in LRP relevance between groups. Future iterations might examine temporal differences in relevance between subtypes [9], [10]. Additionally, it might be helpful to apply a clustering explainability approach to quantify the relative importance of each network pair to each SZ subtype using a modified version of the methods used in existing studies [7], [11].…”