Developing and applying a computationally tractable model to estimate higher-order correlations in fMRI data

Brisbane Convention & Exhibition Centre 

In order to better understand the dynamic interactions between brain structures that underlie our thoughts, the tools and data we use must reflect the complexity of this system. We have developed a computationally tractable model, Timecorr, to estimate higher-order correlations that combines a technique to calculate dynamic correlations with dimensionality reduction. We have applied this model to explore dynamic higher-order correlations in brain data collected using a naturalistic paradigm at varying levels of engagement. In an fMRI dataset collected by Simony et al. (2016), participants listened to a story presented in three conditions: intact, paragraph-scrambled, and word-scrambled. We used a subset of the data to train across-participant classifiers to decode listening times, and we trained the classifiers using iteratively increasing orders of dynamic correlation which we inferred using our method. By training decoding models on different types of neural features, we can better understand which specific aspects of the neural activity patterns are informative, and important for higher order cognition.
We found that classifiers trained to decode from high-order dynamic correlations yield the best performance on data collected as participants listened to the (unscrambled) story. In contrast, classifiers trained to decode data from scrambled versions of the story yielded the best performance when they were trained using first-order dynamic correlations or non-correlational activity patterns. Our results suggest that as our thoughts become more complex, they are reflected in higher-order patterns of dynamic network interactions throughout the brain. We have enhanced and expanded on this work, using other fMRI datasets from the ‘Narratives’ collection Nastase et al. (2020), and diving deeper into the choices of dimensionality reduction and hyperparameters across these datasets. By exploring these parameters across these datasets, we can evaluate the similarities in task-evoked brain responses and explore the level of interactions that support complex thought.