Machine learning and individual variability in electrical field characteristics predict tDCS treatment response for...

Anxiety is highly prevalent in older adults and often co-occurs with neurodegenerative disorders, worsening cognitive decline and quality of life. Transcranial direct current stimulation (tDCS) has shown therapeutic potential, but outcomes remain inconsistent due to individual neurophysiological variability.

We trained machine learning (ML) models to predict state anxiety reduction following active tDCS paired with BrainHQ cognitive training (tDCS + CT) in older adults with moderate/severe baseline state anxiety symptoms from the Augmenting Cognitive Training in Older Adults (the ACT trial, NCT02851511). Our models predicted outcomes based on MRI-derived current density maps (J-maps) generated via finite element modeling. We evaluated the model using (1) repeated mixed-site nested cross-validation and (2) repeated cross-site held-out-site validation; in both settings, analyses were conducted on the 20 eligible participants.

The model achieved a mean balanced accuracy of 82% in the repeated mixed-site nested cross-validation and 72% in the repeated cross-site held-out validation. SHAP-based interpretation of our ML models identified the current density in posterior fusiform cortex and inferior temporal gyrus (posterior and temporooccipital divisions) as the most influential predictors of intervention response.

Posterior ventral temporal cortices emerged as critical substrates for predicting intervention response, consistent with prior studies linking fusiform and amygdala-temporal dynamics to anxiety severity and fearful face processing. These findings highlight how perceptual processes, particularly those mediated by the fusiform gyrus, shapes intervention response. Altogether, this work paved the way for ML-guided precision modeling to personalize non-invasive transcranial electrical interventions for anxiety in older adults.