# Craving Prediction From fMRI Drug Cue Reactivity in Methamphetamine Use Disorder: A Parsimonious Neurobiological Model

## TL;DR

This study uses brain scans and machine learning to predict drug cravings in methamphetamine users, offering a more objective way to assess addiction.

## Contribution

A brain-based model using fMRI and machine learning to predict craving intensity in methamphetamine use disorder is developed and validated.

## Key findings

- The model achieved an RMSE of 0.983 and a mean Pearson correlation of 0.216 in predicting craving intensity.
- Key brain regions like the parahippocampal gyrus and amygdala were positively associated with craving.
- The model successfully classified high versus low craving levels with an AUC-ROC of 0.684.

## Abstract

Craving is a fundamental aspect of substance use disorder (SUD), traditionally assessed through subjective self‐report measures. To develop more objective assessments, we created a brain‐based marker to predict craving based on machine learning approaches using functional magnetic resonance imaging (fMRI) drug cue reactivity data from 69 participants with methamphetamine use disorders.

To predict craving intensity (rated on a 1–4 scale), we developed a modeling pipeline in which multiple feature selection methods (ANOVA, PCA) and regression algorithms (linear regression, Lasso, Elastic Net, Random Forest, and XGBoost) were evaluated. Model performance was assessed using subject‐level 5‐fold cross‐validation plus a 20% hold‐out test set. PCA combined with linear regression yielded the best performance in terms of Root Mean Squared Error (RMSE) while maintaining interpretability. Statistical significance was tested via permutation tests. Model weights were back‐projected to voxels and summarized in the Brainnetome atlas. In addition, the model successfully classified high and low craving levels and distinguished cue types (neutral vs. drug) based on fMRI data.

The model achieved an RMSE of 0.983 ± 0.026 (standard deviation) and a mean Pearson correlation of 0.216, with strong generalization evidenced by an out‐of‐sample RMSE of 0.985 and statistical significance (p < 0.026; effect size (Cliff's Delta) = 0.715; statistical power = 0.639). Key neurobiological signatures included the parahippocampal gyrus, superior temporal gyrus, medioventral occipital cortex, and amygdala (positively associated with craving), as well as the inferior temporal gyrus (negatively associated). Classification of high versus low craving levels yielded an AUC‐ROC of 0.684 ± 0.084 (out‐of‐sample AUC‐ROC = 0.714), with significant separation (p < 0.04; Cliff's Delta = 0.831). In addition, classification of cue types (neutral vs. drug) achieved an AUC‐ROC of 0.692 ± 0.090 (out‐of‐sample AUC‐ROC = 0.693), with p < 0.002, Cliff's Delta = 0.896, and statistical power = 0.800, highlighting the robustness of the model.

These findings underscore the potential of neuroimaging and machine learning to provide objective, data‐driven insights into the neural mechanisms underlying subjective experience of craving and to inform future clinical applications in SUD.

Brain responses to drug cues measured with fMRI, combined with craving self‐reports, were used to train a machine learning model that predicted craving intensity and classified high versus low craving states. These findings highlight the promise of neuroimaging‐based biomarkers for advancing personalized interventions in addiction treatment.