NeuralCipher Research overview
The text below is the source article’s original abstract under a verified reuse license. cc by
Abstract Closed-loop neuromodulation requires a feedback signal that reports therapeutic state, yet such signals are typically read from the same circuits that stimulation is perturbing. A neural feature that changes during stimulation may therefore simply reflect engagement of the circuit rather than clinical improvement. This entanglement problem is common to any stimulation paradigm in which the recording and stimulation share a substrate. Here, we show that the neural signature of stimulation and the signature of therapeutic response can be separated within a single learned representation, without supervision of clinical outcome. We use generative causal representation learning to learn a supervised latent dimension to detect stimulation presence rather than therapeutic state in a low-dose condition, where stimulation engages the circuit, but clinical benefits are minimal. This quarantines the stimulation signature, allowing an orthogonal subspace to carry therapeutic structure. We test the approach in subthalamic deep brain stimulation for Parkinson’s disease, where the confound persists even after the exogenous stimulation artifact is removed. Applied unchanged to recordings at clinically-relevant stimulation amplitudes across four frequencies in nine patients, the stimulation axis classified stimulation state under leave-one-subject-out validation (AUC = 0.919 and 0.942 at half and full amplitude) but did not predict motor outcome, whereas the auxiliary subspace tracked graded motor improvement (ρ = 0.50), survived mixed-effects control for frequency, amplitude and patient identity, and identified low-frequency stimulation as a large anti-therapeutic displacement not captured by beta or gamma power. The frozen representation generalized to a pre-specified held-out patient (rigidity ρ = 0.731) and an independent prospective video-kinematic cohort (n = 3). Beyond DBS, the approach offers a template for discovering biomarkers of therapeutic effect in any closed-loop setting where stimulation contaminates the recording used for feedback, from other neurological and psychiatric DBS targets to spinal cord and non-invasive brain stimulation.
Scientific assessment
Research question
Assess the question within the original study’s scope.
Methods & sample
Check the source article for study design and sample details.
Key findings
An editorially approved findings summary is not yet published for this record.
Limitations
Generalizability, bias and clinical meaning require separate assessment.
Methods, samples and clinical conclusions are not inferred from bibliographic metadata.