Machine learning and individual variability in electric field characteristics predict tDCS treatment response

Abstract

Background: Transcranial direct current stimulation (tDCS) is widely investigated as a therapeutic tool to enhance cognitive function in older adults with and without neurodegenerative disease. Prior research demonstrates that electric current delivery to the brain can vary significantly across individuals. Quantification of this variability could enable person-specific optimization of tDCS outcomes. This pilot study used machine learning and MRI-derived electric field models to predict working memory improvements as a proof of concept for precision cognitive intervention.

Methods: Fourteen healthy older adults received 20 minutes of 2 mA tDCS stimulation (F3/F4) during a two-week cognitive training intervention. Participants performed an N-back working memory task pre-/post-intervention. MRI-derived current models were passed through a linear Support Vector Machine (SVM) learning algorithm to characterize crucial tDCS current components (intensity and direction) that induced working memory improvements in tDCS responders versus non-responders.

Main results: SVM models of tDCS current components had 86% overall accuracy in classifying treatment responders vs. non-responders, with current intensity producing the best overall model differentiating changes in working memory performance. Median current intensity and direction in brain regions near the electrodes were positively related to intervention responses (r=0.811,p<0.001 and r=0.774,p=0.001).

Conclusions: This study provides the first evidence that pattern recognition analyses of MRI-derived tDCS current models can provide individual prognostic classification of tDCS treatment response with 86% accuracy. Individual differences in current intensity and direction play important roles in determining treatment response to tDCS. These findings provide important insights into mechanisms of tDCS response as well as proof of concept for future precision dosing models of tDCS intervention.

Keywords: Cognitive aging; Finite element modeling; Machine learning; Transcranial direct current stimulation; Treatment response; tDCS.

Fig. 1.

An example of the stimulus presentation in the two-back variant of the N-back task.

Fig. 2.

A) The F3-F4 electrode montage with the F3 electrode as the cathode (blue) and the F4 electrode as the anode (red). B) A representative image of actual anode (red) and cathode (blue) placements for a single subject across ten sessions. The mean displacement for this participant of the anode and cathode was 1.9 cm (−0.2 St. Dev.) and 1.7 cm (−0.45 St. Dev.), respectively.

Fig. 3.

A) Areas under the ROC curve across ten iterations of three data types: direction, intensity, and combined. Mean AUC plotted as bars. No significant difference was observed. B) The mean ROC curve of each model across ten iterations with shaded area conveying 95% confidence interval.

Fig. 4.

Discrimination maps of regions that predict working memory improvements with the percent contribution of each voxel to the SVM decision function, superimposed onto the MNI152 Template.

Fig. 5.

Plots to demonstrate the current density characteristics within regions predictive of tDCS responders. A) Histogram of current intensity (bin width of 0.0013 Am-2) with the y-axis representing the number of observations in each bin divided by the total number of observations, where the sum of all bar heights is equal to 1. B) Cumulative histogram of current intensity with the height of each step equal to the cumulative number of observations in the bin over the total number of observations in each bin and all previous bins where the height of the last bar is equal to 1. C) Scatter plot of behavior change (post – pre intervention working memory performance) vs. median current intensity. D) The Hedges’ g between responders and non-responders is shown in a Gardner-Altman estimation plot. The mean difference is plotted on a floating axis as a bootstrap sampling distribution. The mean difference is depicted as a dot; the 95% confidence interval is indicated by the ends of the vertical error bar.

Fig. 6.

Single representative participant A) Coronal, B) Sagittal, and C) Axial image of current intensity represented by the color of the images, and current direction represented by the arrows within the images. The color bar represents electric field in volts per meter (v/m). D) Histogram of azimuthal angle δ with the height of each bar representing the number of observations in each bin divided by the total number of observations, where the sum of the all bar heights is equal to 1.

Fig. 7.

A) Visualization of the top 10 regions of interest from the Harvard-Oxford atlas ranked based on their contribution toward predictions of treatment response. B) Rank, label, and mean percent contribution per voxel of the top ten regions of interest. C) a bar graph to represent the average percent contribution per voxel within each ROI of Harvard-Oxford atlas with the top 10 regions of interest highlighted in red.

Fig. 8.

A) Anode and B) Cathode displacement between responders and non-responders, with means plotted as bars. Non-responders were found to have greater displacement of both electrodes (anode: F 2 = 6.73, P = 0.023, cathode: F 2 = 19.39, p < 0.001) from their ideal placement compared to responders. Linear regression of C) current intensity and D) current direction, based on the percent difference of average current intensity and cathode displacement versus behavioral change. The optimal stimulation parameters are represented by the diamond in the figure (i.e., the mean within the responder group). * represents p < 0.05 and ** represents p < 0.01).