(J-364) Non-Linear Methods for Decoding High Risk Decision Making

Introduction:: Decision-making is a complex cognitive phenomenon that requires the coordinated activity of multiple brain areas and encompasses multiple sub-processes, including reward evaluation, option comparison, and choice selection. Theories surrounding decision-making generally posit a form of information accumulation followed by choice, which has been supported by previous neurological and computational studies. However, disentangling the precise contribution of individual human brain areas over time to decision-making remains challenging, primarily due to the constraints imposed by non-invasive neural recording methods. These constraints limit the efficacy of modern nonlinear computational models, which require large amounts of high-quality data to perform well. Despite research indicating that the brain encodes information non-linearly, most recent work has focused on linear models


Materials and Methods:: To address this methodological impediment, our team conducted a study on a cohort of twenty epilepsy patients undergoing intracranial electrophysiological monitoring. In this experiment, we recorded neural activity from several brain regions involved in reward processing, including the prefrontal and parietal cortices, as well as deep temporal lobe structures. While being monitored, the participants engaged in a risky decision-making task that required them to choose between a safe bet and a risky gamble for a higher monetary reward. Here, we modeled our data on a patient-by-patient basis using non-linear methods that maintain interpretability while obtaining high classification accuracy of whether the participant would gamble or not. Specifically, we employed 1-D Convolutional Neural Networks analyzed with Class Activation Maps and Random Forests analyzed with Gini Importance.


Results, Conclusions, and Discussions:: Our models both achieved a higher classification accuracy than previous linear models(79.7± 3.63%) and allowed greater insight into the accumulation of information prior to choice. We found from analyzing the class activation maps that, over time prior to choice, the brain alternates between encoding information about the gamble choice and the safe choice up until ~200ms before the choice is made. This occurs at roughly 2-5 Hz, and the timing is consistent for each subject. This alternating pattern of information encoding may be linked to an individual's delta waves, which have also been shown to be linked to attention. These results support the accumulation of information hypothesis but also indicate that the brain accumulates information specific to each potential decision separately. The Gini Importance highlighted that the neural information about choice was distributed throughout the brain and combined with ablation studies that removed each brain region from the analysis without significant loss in classification accuracy, suggests that information about choice is distributed across the brain with some redundancy. Future studies will examine the interaction between different brain regions during the accumulation of information about each choice and the implications for each region's function in the overall decision-making process.


Acknowledgements (Optional): :

References (Optional): :