Objective: To develop an information-theoretic framework that identifies neural signatures predicting therapeutic benefit, independent of stimulation artifacts, enabling objective DBS parameter optimization.

Background: DBS programming requires extensive trial-and-error titration of parameters. Even closed-loop strategies rely on empirically selected feedback signals (e.g., beta power) that themselves require tuning. LFP recordings during stimulation contain both therapeutic responses and artifacts, but current methods cannot separate features predicting clinical benefit from those detecting stimulation presence, perpetuating empirical adjustment.

Method: We analyzed bilateral LFP recordings from 9 Parkinson’s patients during ON stimulation (125, 110, 85, 55 Hz at half and full amplitude) and OFF, extracting 70 features (spectral power, connectivity, nonlinear dynamics, phase-amplitude coupling). Generative Causal Explanations partitioned a variational autoencoder latent space into K∈{1:3} stimulation residual components (SRC; maximizing mutual information with an ON/OFF classifier) and L∈{1:6} physiological latent components (PLC; independent of classifier outputs). Models trained on half-amplitude were validated on full-amplitude recordings. Latent representations were correlated with normalized UPDRS III scores and therapeutic axes confirmed by linear mixed-effects models (LME) with random patient intercepts.

Results: Optimal configuration was K=1, L=5. The SRC classified ON vs OFF stimulation with leave-one-subject-out AUC=0.919 [95% CI: 0.831–0.981] yet showed no clinical correlation (ρ=−0.032, p=0.85), confirming detection–therapy separation. On held-out full-amplitude recordings (model trained on half-amplitude), signed PLC1–PLC2 displacement correlated with total motor improvement (ρ=0.497, p=0.002; n=36). LME with random patient intercepts and frequency as covariate confirmed a joint PLC1–PLC2 therapeutic axis (β=−0.776, p=0.008; permutation p=0.040); PLCs 3–5 were null (all p>0.16).

Conclusion: Information-theoretic partitioning separates therapeutic mechanisms from artifacts, revealing amplitude-invariant physiological biomarkers. This framework enables objective biomarker identification during active stimulation, potentially replacing empirical titration with data-driven optimization.

References: 1. O’Shaughnessy, M., Canal, G., Connor, M., Rozell, C. & Davenport, M. Generative causal explanations of black-box classifiers. Adv. Neural Inf. Process. Syst. 33, 5453–5467 (2020).
2. Alagapan, S., Choi, K.S., Heisig, S. et al. Cingulate dynamics track depression recovery with deep brain stimulation. Nature 622, 130–138 (2023). https://doi.org/10.1038/s41586-023-06541-3

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/delineating-therapeutic-and-non-therapeutic-neural-signatures-of-deep-brain-stimulation/