Classifying impulsivity in cocaine use disorder from fMRI brain maps

Introduction: : Delay discounting - the tendency to devalue future rewards in favor of immediate ones – is a behavioral marker of addiction and impulse disorders. In alcohol use disorder (AUD), higher discounting rates (k-values) have been linked to greater alcohol dependence, entry into alcohol use, quantity of use, addiction severity, and poorer treatment outcomes (Bickel et al., 2014). This study investigated whether functional magnetic resonance imaging (fMRI) findings from (AUD) generalize to cocaine use disorder (CUD) in the context of delay discounting.
We used an fMRI-based delay discounting task that could generate multiple brain maps. Specifically, we examined how the brain responds to decisions regarding taking a smaller amount of money sooner vs. taking a larger amount of money later. By sorting each decision trial into different contrasts, we produced brain maps related to "Hard vs. Easy" trials, "Immediate Available vs. Immediate Not Available" trials (examining the extra cognitive load elicited when a “now” option was not presented in the trial), and “All Trials,” (irrespective of their categories). With each of these three maps from every CUD participant, we trained “high” vs. “low“ discounter classifiers. Comparing model performance across maps identifies which trials most heavily encode impulsivity information, and what brain regions are most useful for classifying an individual’s k-value. Identifying these predictive regions may provide mechanistic insight into addiction vulnerability.

Materials and

Methods: : Forty-six participants with cocaine use disorder were included in this study. fMRI data was acquired at 3 T. Delay discounting (DD) k-values were derived from an in-scanner individualized delay discounting task. Preprocessing of the fMRI data was performed using afni_proc.py, with the following specifications slice timing correction, aligned to minimum outlier volume, non-linear warp to common space, blur to 6mm FWHM, scale to %signal change. Three contrasts were generated using subject-level general linear model (GLM) analysis to capture decision-making behavior: 1) Hard vs Easy; 2) Immediate Available vs Immediate Not Available; and 3) the All-Trials contrast. Subject-level contrast maps were concatenated into three group-level files using 3dTcat. Masking was performed using the afni 3dmask_tool on a general mask with the -dilation_input -2 option selected. In-scanner k-values were used to group participants into low, medium, and high tertiles. A two-class support vector classifier (SVC) was implemented using 3dSVM to classify participants as high vs low discounters. The classifier performance was evaluated using Leave-One-Pair-Out cross-validation. Post-classification, SVM weight maps were examined to identify voxel clusters that contributed most to group classification.

Results, Conclusions, and Discussions:: Results
Classifier accuracy results are shown in Fig. 1. Across all classification methods, performance varied notably by contrast type. The Hard vs. Easy contrast achieved the best overall performance – 74.2 % classification accuracy using Leave-One-Pair-Out– as well as higher general performance. The Immediate Available vs. Immediate Not Available contrast performed worse than chance, and the All-Trials contrast was slightly above chance. These results stress the importance of identifying relevant contrasts and implementing them in task design for accurate predictions of delay discounting behavior.
Support vector machine (SVM) weight maps were analyzed to identify key regions that contributed to predicting high and low discounters using the Hard vs. Easy contrast maps. The SVM model relied heavily on regions including the caudate, inferior temporal, rostral anterior cingulate, bilateral insula, and posterior cingulate to predict high or low discounting (Fig. 2).

Conclusion
In this study, we developed a two-class support vector classifier SVC to differentiate participants with high versus low delay discounting behaviors based on individual fMRI activation maps of relevant task contrasts. This classifier achieved significant accuracy above chance, demonstrating that these task-specific contrast maps contain relevant neural correlates for predicting discounting behaviors. The brain regions identified - caudate, inferior temporal, rostral anterior cingulate, bilateral insula, and posterior cingulate – are consistent with previous research linking these areas to delay discounting. These contrast-specific brain activation maps could serve as a potential biomarker for addiction vulnerability, potentially providing a useful tool for clinical applications.

Acknowledgements and/or References (Optional):: Bickel WK, Koffarnus MN, Moody L, Wilson AG. The behavioral-and neuroeconomic process of temporal discounting: a candidate behavioral marker of addiction. Neuropharmacology. 2014; 76:518–527. doi: 10.1016/j.neuropharm.2013.06.013

Amlung M, Vedelago L, Acker J, Balodis I, MacKillop J. Steep delay discounting and addictive behavior: a meta-analysis of continuous associations. Addiction. 2016; 112(1):51-62. doi: 10.1111/add.13535

Koffarnus MN, Deshpande HU, Lisinski JM, Eklund A, Bickel WK, LaConte SM. An adaptive, individualized fMRI delay discounting procedure to increase flexibility and optimize scanner time. Neuroimage. 2017 Nov 1;161:56-66. doi: 10.1016/j.neuroimage.2017.08.024. Epub 2017 Aug 10. PMID: 28803942; PMCID: PMC5696088.

Cox RW. AFNI: software for analysis and visualization of functional magnetic resonance neuroimages. Comput. Biomed. Res. 1996; 29(3):162–173. doi: 10.1006/cbmr.1996.0014

LaConte S, Strother S, Cherkassky V, Anderson J, Hu X. Support vector machines for temporal classification of block design fMRI data. NeuroImage. 2005; 26(2):317-329. doi: 10.1016/j.neuroimage.2004.08.034

Kable JW, Glimcher PW. The neural correlates of subjective value during intertemporal choice. Nature Neuroscience. 2007; 10(12):1625–1633. doi: 10.1038/nn2007

Hare TA, Hakimi S, Rangel A. Activity in dlPFC and its effective connectivity to vmPFC are associated with temporal discounting. Frontiers in Neuroscience. 2014; 8:50. doi: 10.3389/fnins.2014.00050

McClure SM, Laibson DI, Loewenstein G, Cohen JD. Separate neural systems value immediate and delayed monetary rewards. Science. 2004; 306(5695):503-507. doi: 10.1126/science.1100907