Predicting Neuromodulation Outcome for Parkinson's Disease with Generative Virtual Brain Model

Abstract

Parkinson's disease (PD) affects over ten million people worldwide. Although temporal interference (TI) and deep brain stimulation (DBS) are promising therapies, inter-individual variability limits empirical treatment selection, increasing non-negligible surgical risk and cost. Previous explorations either resort to limited statistical biomarkers that are insufficient to characterize variability, or employ AI-driven methods which is prone to overfitting and opacity. We bridge this gap with a pretraining-finetuning framework to predict outcomes directly from resting-state fMRI. Critically, a generative virtual brain foundation model, pretrained on a collective dataset (2707 subjects, 5621 sessions) to capture universal disorder patterns, was finetuned on PD cohorts receiving TI (n=51) or DBS (n=55) to yield individualized virtual brains with high fidelity to empirical functional connectivity (r=0.935). By constructing counterfactual estimations between pathological and healthy neural states within these personalized models, we predicted clinical responses (TI: AUPR=0.853; DBS: AUPR=0.915), substantially outperforming baselines. External and prospective validations (n=14, n=11) highlight the feasibility of clinical translation. Moreover, our framework provides state-dependent regional patterns linked to response, offering hypothesis-generating mechanistic insights.

Figures (6)

• Figure 1: Clinical gap, proposed framework, and cohort statistics.a, Overview of trial-and-error neuromodulation process for PD and the clinical situation. b, Overview of the proposed framework: a generative virtual brain paradigm transfers dynamical priors from large-scale data to small clinical cohorts for individualized neuromodulation response prediction. c--d,Statistics of the pretraining cohort: sample distribution across four centers (c) and disease spectrum composition (d). e--g,Statistics of the Ruijin PD cohorts: age distribution stratified by sex (e), MDS-UPDRS III motor scores distribution (f) and PD severity composition (g). h, Image processing pipeline: from raw BOLD acquisition to parcellated regional time series.
• Figure 1: Full Heatmap of absolute Spearman correlation strengths between brain regions and MDS-UPDRS symptom measures. Rows denote individual symptoms. Columns show the top-ranked regions, defined by the number of associations exceeding the threshold ($|\rho| \ge 0.3$ and $p < 0.05$). Color represents $|\rho|$, with darker shades indicating stronger effects. Associations not meeting the threshold or not confirmed by the significance table are displayed in gray. Results are presented for TI and DBS cohorts separately.
• Figure 2: Foundation and individualized virtual brain validation.a, Architecture of the generative virtual brain model. b, BOLD signal forecasting in four Parkinson's disease–relevant regions on the test set and global MAE across 166 regions: ANN = 0.7044, FVB = 0.6576, iVB = 0.5346. c, Per-region R2 distribution across 166 brain areas. d, Cumulative distribution function (CDF) of R2 values. e--g, Functional connectivity validation for ANN (e; $r = 0.630$), FVB (f; $r = 0.867$), and iVB (g; $r = 0.948$) against empirical data (all $p < 10^{-200}$, two-tailed Pearson correlation). h, Visualization of brain functional connectivity using Yeo 7-network parcellation illustrates that iVB preserves high-level connectivity patterns closely matching empirical data. i, Boxplots w.r.t. the performance comparison between FVB and iVB across 106 patients (TI and DBS cohorts) show the significant improvements in MAE, R2, and FC correlation after finetuning (all $p < 0.001$, two-tailed paired t-test).
• Figure 3: Validation of predicting neuromodulation response.a, Schematic illustration of the iVB-based workflow for predicting neuromodulation response. b, Diagram depicting the calculation of the Counterfactual Brain Mismatch. c, Clinical variable comparisons between responders/non-responders in TI ($n = 51$) and DBS ($n = 55$) cohorts. Fisher's exact test for sex (Male/Female) and target (GPi/STN); Mann–Whitney U test for age. All comparisons non-significant (ns; $p>0.05$). d--f, Five-fold cross-validation performance comparing the proposed iVB-based model (Ours) with baselines. d, Summary metrics (mean $\pm$ 95% CI) for Accuracy (ACC), F1, and Precision. Ours achieved strong performance in both TI (ACC = 0.845, F1 = 0.816, Precision = 0.843) and DBS (ACC = 0.836, F1 = 0.847, Precision = 0.917), significantly outperforming all baselines (paired $t$-tests, all $p < 0.05$). Curves show mean ROC and PR with 95% CI. e, TI cohort: AUC = $0.823 \pm 0.083$, AUPR = $0.853 \pm 0.066$ (ours). f, DBS cohort: AUC = $0.865 \pm 0.072$, AUPR = $0.915 \pm 0.048$ (ours). g--h, Forest plots report subgroup-wise AUC (circles) and AUPR (squares) across five cross-validation folds for the TI (g) and DBS (h). Logistic interaction tests showed no significant effect modification (all $p>0.05$)
• Figure 4: Clinical utility and validation of the iVB-based prediction framework.a, Decision curve analysis (DCA) for the proposed model in TI and DBS cohorts. Our model (solid lines) outperforms reference strategies "Treat all" (dashed blue) and "Treat none" (dashed yellow). b, Raincloud plots of predicted response probabilities stratified by clinical response status (Responders vs. Non-responders). Statistical comparisons: TI cohort (Mann-Whitney $U=33.00$, $p<0.001$, Cohen's $d=2.39$); DBS cohort (Mann-Whitney $U=132.00$, $p<0.001$, Cohen's $d=1.29$). c--d, Comparison of individual functional connectivity correlations versus integrated model prediction for TI (c) and DBS (d). The first four panels show Spearman correlations between clinical improvement and connectivity measures from different brain regions (left STN-GPi, left STN-Thalamus, STN-M1, STN-SMA). The fifth panel displays the iVB-based model results pooled from a five-fold cross-validation test data, which achieves significantly stronger correlation with treatment outcomes compared to individual connectivity measures. e--f, Model prediction ROC, and Precision-Recall curves, prediction spectrum plots, for external (e, n = 14) and prospective (f, n = 11) validation. Curves show mean ± std over five random seeds, with BrainLM and BrainGNN as baselines.