A Whole-Brain Connectome-Constrained Model of Subcallosal Cingulate Deep Brain Stimulation Evoked Potentials
DOI: Pending DOI
Synopsis
Deep brain stimulation (DBS) of the subcallosal cingulate cortex (SCC) is an emerging intervention for treatment-resistant depression. SCC stimulation-evoked potentials (SEPs) show reproducible EEG responses, indicating consistent engagement of DBS-induced circuits, but how specific pathways and circuit dynamics generate these signals remains unclear. In this study, we present a whole-brain, connectome-constrained neural model with machine learning-based parameter estimation to link structural and neural mechanisms underlying SCC DBS to SEPs. Fitted to high-density EEG, the model reproduced the spatiotemporal SEP patterns and generalized to unseen stimulation amplitudes. This framework provides a mechanistic tool for interpreting SEP biomarkers and probing DBS-induced circuit dynamics.
Background
Deep brain stimulation (DBS) of the subcallosal cingulate cortex (SCC) has emerged as a promising intervention for treatment-resistant depression (TRD) [1]. The electrophysiological effects of SCC DBS evolve non-linearly across timescales [1,2], suggesting an adaptive circuit-level process. The neural mechanisms of SCC DBS are now emerging, but not yet fully understood. Given the limitations of current experimental techniques for probing human neural circuits in vivo, electrophysiological readouts, particularly scalp electroencephalography (EEG), provide a non-invasive and scalable approach to study DBS-induced neural dynamics. Stimulation-evoked potentials (SEPs) capture time-locked cortical responses that may reflect the propagation of DBS-evoked activity along white matter pathways and within distributed networks. Prior work has shown that SCC SEPs are reliably reproduced across patients and sessions, indicating a consistent engagement of DBS-induced circuits [3], and that longitudinal changes in SEPs are associated with clinical response in chronic DBS [4]. However, SEP-based biomarkers remain largely descriptive and does not reveal how engagement of specific pathways contribute the observed signals. Computational approaches integrating structurally constrained biophysical models with empirical SEPs can help infer DBS-induced effects on engaged circuits. Here, we present a computational framework that combines whole-brain connectome-constrained modeling with machine learning-based parameter estimation to mechanistically link SCC stimulation to scalp-recorded SEPs.
Methods
High‑density scalp EEG was recorded during SCC SEPs in individuals with TRD. SEPs were defined as cortical responses time‑locked to 2 Hz DBS pulses at 6V with coherent topography on the head surface [3]. SEP data were windowed around each pulse and preprocessed to remove stimulation artifacts using independent component analysis. Subject‑level evoked response was obtained by averaging artifact‑cleaned trials. A whole‑brain connectome‑constrained model of neural dynamics was developed by leveraging the structural connectivity data from the Human Connectome Project [5]. Each brain region was modeled with an extended Jansen‑Rit neural mass model capturing three neural population including the pyramidal neurons, excitatory stellate cells, and inhibitory interneuron subpopulations [6]. SCC stimulation was represented as an external input to the appropriate node. A forward head model was then used to project simulated cortical activity onto 256 scalp channels, generating synthetic DBS‑evoked potentials. The model was implemented in PyTorch, enabling efficient gradient‑based optimization via automatic differentiation [7, 8]. Model parameters were estimated by minimizing the mean squared error between simulated and experimental SEPs. Training was performed over multiple runs and initializations to assess convergence and robustness of the inferred parameters.
Results
We trained the model using averaged stimulation-induced EP data from TRD patients undergoing SCC DBS. Model training proceeded for 120 epochs, during which the mean squared error between simulated and experimental evoked potentials decreased and plateaued after approximately 100 epochs, indicating convergence of the parameter estimates. To assess how well the trained model captured large‑scale network dynamics, functional connectivity matrices were computed from both simulated and experimental SEP time series. The resulting matrices showed high similarity, with an average correlation coefficient of 0.84 ± 0.01 across five independent training runs. In addition, the spatiotemporal patterns of simulated stimulation‑evoked potentials closely matched the experimental data, with a cosine similarity of 0.85 ± 0.03 between simulated and recorded sensor‑level waveforms. The model also preserved characteristic topographic patterns, suggesting that it captured stable network pathways mediating SCC stimulation effects. These findings indicate that the connectome‑constrained whole brain model can reproduce both the temporal dynamics and spatial distribution of SCC SEP responses. Generalizability was examined by probing the trained model with stimulation amplitudes that was not used during training. Stimulation amplitude was changed from 1 to 6V in the whole‑brain network model while holding all learned parameters fixed. The simulated SEPs showed systematic amplitude‑dependent changes, including increased SEP amplitude and wider waveforms with higher stimulation intensities, resembling the observed changes in amplitude sweep experimental SEP data. Together, these results demonstrate that the proposed framework can fit empirical SCC SEPs, recover functional connectivity and spatiotemporal patterns, and generalize to unseen stimulation conditions, supporting its use as a mechanistic framework for investigating DBS induced circuit dynamics.
Discussion
This work presents a biophysically grounded, connectome‑constrained modeling framework that quantitatively links SCC DBS to scalp‑recorded SEPs in patients with treatment‑resistant depression. By integrating the extended Jansen‑Rit neural mass models in a whole‑brain connectome-constrained model and fitting them directly to high‑density EEG recordings, the framework captures the spatial and temporal features of SCC SEPs and reproduces functional connectivity patterns inferred from empirical data. The ability of the trained model to generalize across stimulation amplitudes further supports that it captures stable and physiologically meaningful circuit properties.
From a clinical and translational perspective, this approach moves beyond descriptive SEP biomarkers by providing a mechanistic tool to quantify and infer how specific pathways and local microcircuit parameters contribute to observed SEP signatures of DBS. The framework enables systematic in silico perturbations, such as selectively altering connection strengths or local excitatory‑inhibitory balance, to test hypotheses about how SCC stimulation modulates communication within and between networks. In the future, such models could be used to explore network engagement of DBS-induced circuits and interpret the neural mechanism underlying longitudinal changes in SEP features.
Disclosures
Research supported by UH3NS103550, UH3NS141080, Georgia CTSA TL1 (TL1TR002382), UL1 (UL1TR002378), and KL2 (KL2TR002381) grants.
References
1. Alagapan S, Choi KS, Heisig S, Riva-Posse P, Crowell A, Tiruvadi V, Obatusin M, Veerakumar A, Waters AC, Gross RE, Quinn S. Cingulate dynamics track depression recovery with deep brain stimulation. Nature. 2023 Oct 5;622(7981):130-8.
2. Cha J, Choi KS, Rajendra JK, McGrath CL, Riva-Posse P, Holtzheimer PE, Figee M, Kopell BH, Mayberg HS. Whole brain network effects of subcallosal cingulate deep brain stimulation for treatment-resistant depression. Molecular Psychiatry. 2024 Jan;29(1):112-20.
3. Waters AC, Veerakumar A, Choi KS, Howell B, Tiruvadi V, Bijanki KR, Crowell A, Riva‐Posse P, Mayberg HS. Test–retest reliability of a stimulation‐locked evoked response to deep brain stimulation in subcallosal cingulate for treatment resistant depression. Human brain mapping. 2018 Dec;39(12):4844-56.
4. Desai A, Van Bogaert T, Choi KS, Song HN, Cha J, Alagapan S, Xu E, Dahill-Fuchel J, Veerakumar A, Kopell BH, Figee M. Longitudinal Changes in Cortical Response Dynamics with Deep Brain Stimulation to the Subcallosal Cingulate for Treatment-Resistant Depression. bioRxiv. 2025 Dec 1:2025-11.
5. Van Essen DC, Ugurbil K, Auerbach E, Barch D, Behrens TE, Bucholz R, Chang A, Chen L, Corbetta M, Curtiss SW, Della Penna S. The Human Connectome Project: a data acquisition perspective. Neuroimage. 2012 Oct 1;62(4):2222-31.
6. Sanz-Leon P, Knock SA, Spiegler A, Jirsa VK. Mathematical framework for large-scale brain network modeling in The Virtual Brain. Neuroimage. 2015 May 1;111:385-430.
7. Paszke A, Gross S, Chintala S, Chanan G, Yang E, DeVito Z, Lin Z, Desmaison A, Antiga L, Lerer A. Automatic differentiation in pytorch.
8. Momi D, Wang Z, Griffiths JD. TMS-evoked responses are driven by recurrent large-scale network dynamics. Elife. 2023 Apr 21;12:e83232.