Individualized functional brain mapping machine learning prediction of symptom-change resulting from selective...

Anhedonia remains a difficult-to-treat symptom and has been associated with poor clinical course transdiagnostically. Here, we applied machine learning models to individualized neural patches derived from fMRI data during the Monetary Incentive Delay Task in anhedonic participants (N = 67) recruited for a clinical trial examining K-opioid receptor (KOR) antagonism in the treatment of anhedonia.

Nine ensemble models were estimated using cortical, subcortical, and combined cortical subcortical features from individualized functional topographies to predict changes in symptoms of overall psychopathology (anhedonia, depression, anxiety). Analyses were performed on the KOR (N = 33) and placebo (N = 34) group.

Initial models showed that only subcortical data predicting depression and anxiety symptom change had a significant Spearman correlation between veridical and predicted data (rho = 0.480 and rho = 0.415 respectively). Next, leave-one-out-cross-validation (LOOCV) showed that the best-performing models comprised only the subcortical individualized systems data, which correlated with clinical change for depression and anxiety scores for the KOR group with significantly higher accuracy (rho = 0.634 and rho = 0.562, respectively) compared to the placebo group (rho = 0.294 and rho = 0.034, respectively). Further, 25 subcortical neural features were identified based on correlation and ensemble determined importance in driving prediction. Final models for both depression and anxiety showed an overall higher representation of the dorsal attention network. Cortical and combined cortical-subcortical feature data showed no significant improvement in prediction of clinical change between the two groups.

Using an ensemble of machine learning approaches, we identified individual differences in subcortical individualized systems data that predicted clinical change that was specific to KOR antagonism.