🧬 biology

A stimulation-orthogonal neural biomarker of therapeutic response in deep brain stimulation

This paper introduces a generative causal representation learning method that successfully separates stimulation artifacts from therapeutic signals in Parkinson's disease deep brain stimulation, enabling the identification of a stimulation-orthogonal neural biomarker that accurately tracks motor improvement without supervision of clinical outcomes.

Original authors: Tanmoy Sil, Pavel Navratil, Ghadir Abbas, Florian Lange, Jens Volkmann, Muthuraman Muthuraman, Martin Reich, Robert Peach

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tanmoy Sil, Pavel Navratil, Ghadir Abbas, Florian Lange, Jens Volkmann, Muthuraman Muthuraman, Martin Reich, Robert Peach

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to tune a radio to find a clear song, but the radio itself is making a loud, staticky buzzing noise that drowns out the music. In the world of treating Parkinson's disease with Deep Brain Stimulation (DBS), doctors face a similar problem. They use electrical pulses to calm a specific part of the brain, but they also need to "listen" to that same part of the brain to see if the treatment is working.

The problem is that the "listening" signal is contaminated by the "talking" signal. When the machine sends electricity, the brain's reaction looks like a mix of "the machine is on" and "the patient is feeling better." It's hard to tell which is which. If the doctor's computer only sees the "machine is on" signal, it might think the treatment is working when it's actually just doing its job of buzzing, or worse, it might miss the real improvements.

The "Noise-Canceling" Solution

The researchers in this paper came up with a clever way to separate the "buzz" from the "music" using a special kind of artificial intelligence called Generative Causal Explanations (GCE).

Think of their method like a noise-canceling headphone for brain data, but instead of just silencing the noise, it sorts the sound into two different rooms:

Room A (The "Is the Machine On?" Room): This room is trained to only listen for the loud, obvious buzz of the electrical stimulation. It learns to say, "Yes, the machine is on," or "No, it's off."
Room B (The "Is the Patient Better?" Room): This room is forced to ignore the buzz entirely. It can only hear the subtle changes in the brain that happen besides the buzz.

How They Trained the AI

To teach the AI how to do this, the researchers didn't just ask it to guess if a patient felt better (which is hard to measure perfectly). Instead, they used a trick:

They turned the stimulation on at a very low volume (half-strength). At this level, the brain definitely feels the "buzz" (the machine is on), but the patient doesn't feel much relief yet.
They taught the AI to put all the "buzz" information into Room A.
Because the AI was forced to put all the "machine on" data into Room A, Room B was left with only the subtle, leftover brain patterns.

Then, they turned the stimulation up to the full, therapeutic volume. They didn't retrain the AI; they just used the same "frozen" settings.

What They Found

When they looked at the data from the full-strength treatment, the results were like magic:

Room A (The Buzz Detector): It worked perfectly. It could tell with high accuracy whether the machine was on or off, no matter the patient or the volume. But, crucially, it had no idea if the patient's symptoms were improving. It was just a "machine on/off" switch.
Room B (The Therapeutic Detector): This room held the secret. The patterns in this room moved in a way that perfectly matched how much the patients' symptoms (like stiffness and slowness) got better.
- When the stimulation was at the right frequency (like 125 Hz), Room B showed a "good" pattern.
- When the stimulation was at a low, ineffective frequency (55 Hz), Room B showed a "bad" pattern, even though the machine was buzzing just as loudly.

The "Unseen" Test

To prove this wasn't just a lucky guess, they tested the system on two new groups they had never seen before:

One new patient: The system correctly identified the "buzz" and predicted that this specific patient's stiffness improved, just like it did for the others.
A group of three patients filmed by video: Instead of asking doctors to rate the patients, they used video to track finger tapping and hand speed. The "Room B" signal still matched the improvements seen in the video, proving the system works even when you measure movement differently.

The Big Picture

The paper claims that they have found a way to quarantine the signal of the electrical stimulation so that the true signal of therapeutic benefit can be seen clearly.

The Old Way: Trying to guess if the patient is better by looking at the raw brain signal, which is messy because the electricity is in the way.
The New Way: Using a smart filter to separate the "electricity noise" from the "healing signal."

The researchers say this approach could be a template for other types of brain stimulation (like for depression or spinal cord issues) where the recording and the stimulation happen in the same place. They found that by separating the "presence of the tool" from the "effect of the tool," they can finally see the true therapeutic effect clearly.

Technical Summary: A Stimulation-Orthogonal Neural Biomarker of Therapeutic Response in Deep Brain Stimulation

Problem Statement
Closed-loop neuromodulation, particularly in deep brain stimulation (DBS) for Parkinson's disease, faces a fundamental "entanglement problem." In current systems, the neural readout used for feedback is recorded from the same circuit that the stimulation perturbs. Consequently, neural features that change during stimulation may reflect the mere presence of the intervention (circuit engagement) rather than genuine therapeutic improvement. Standard biomarkers, such as beta-band power in the subthalamic nucleus (STN), are highly sensitive to stimulation presence but fail to distinguish between therapeutic and non-therapeutic states. For instance, low-frequency stimulation can engage the circuit without providing clinical benefit, yet standard decoders might misinterpret this engagement as a therapeutic signal. The core challenge is to separate the neural dimensions identifying the presence of stimulation from those measuring the therapeutic effect of stimulation within a single representation, without relying on supervised clinical outcome labels during the separation process.

Methodology
The authors propose a solution using Generative Causal Explanations (GCE), a deep learning framework that partitions high-dimensional neural features into a supervised latent component and orthogonal auxiliary components.

Experimental Design: The study utilized subthalamic local field potentials (LFPs) from nine patients with advanced Parkinson's disease (development cohort) and one prospectively held-out patient (P11). Recordings were acquired at four stimulation frequencies (125, 110, 85, 55 Hz) and two amplitudes (clinically programmed full amplitude and half amplitude).
Training Strategy (The "Quarantine" Approach): Instead of training the supervised dimension to predict clinical outcome (which is confounded by stimulation presence), the authors trained the GCE model to detect stimulation presence (ON vs. OFF) using only half-amplitude recordings. At half-amplitude, stimulation reliably engages the circuit, but clinical benefits are minimal and inconsistent.
- Objective: Force the classifier-aligned latent dimension (Dim 0) to absorb the "stimulation signature."
- Constraint: The remaining auxiliary dimensions (Dims 1–5) are constrained to preserve neural structure orthogonal to the stimulation detection task.
Feature Space: The input consisted of 70 features per 10-second segment, including spectral power (alpha, beta, gamma bands), burst dynamics, phase-amplitude coupling, interhemispheric partial directed coherence (PDC), and nonlinear dynamical measures.
Evaluation: The trained encoder was frozen and applied without retraining to:
1. Full-amplitude recordings across all frequencies in the development cohort.
2. The prospectively held-out patient (P11).
3. An independent cohort of three patients assessed via video kinematics (different protocol, different outcome modality).
Validation: The stimulation-aligned dimension was tested for its ability to classify ON/OFF states (generalizability) and its lack of correlation with motor improvement (negative control). The auxiliary subspace was tested for its ability to track graded motor improvement and distinguish between effective and ineffective stimulation settings.

Key Results

Successful Disentanglement: The stimulation-aligned dimension (Dim 0) successfully generalized as a marker of stimulation presence across patients and amplitudes (Leave-One-Subject-Out AUC = 0.919 at half-amplitude; 0.942 at full-amplitude). Crucially, Dim 0 showed no association with motor improvement (total score, rigidity, bradykinesia, or tremor), confirming it captured only the intervention signature.
Therapeutic Signal in Orthogonal Space: The auxiliary subspace (specifically the combined signed distance of Dims 1 and 2) tracked graded motor improvement at full amplitude (Spearman $\rho = 0.50$ $ρ = 0.50$ , $p=0.002$ $p = 0.002$ ).
- Symptom Specificity: The signal was strongest for rigidity ( $\rho = 0.479$ ) and showed a weaker, readout-dependent association with bradykinesia. It showed no robust relationship with tremor.
- Frequency Dependence: The auxiliary metric identified low-frequency (55 Hz) stimulation as a large "anti-therapeutic" displacement. Despite 55 Hz engaging the circuit strongly (large Euclidean displacement in the auxiliary space), it moved all patients toward the least therapeutic region of the latent space, a distinction not captured by standard beta or gamma power.
Generalization:
- Prospective Patient: In the held-out patient (P11), the auxiliary metric correlated significantly with rigidity improvement ( $\rho = 0.731$ , $p=0.040$ ) without retraining.
- Independent Kinematic Cohort: The model generalized to a separate cohort assessed via markerless video kinematics. The auxiliary metric correlated with finger-tapping consistency and hand-movement speed, validating the therapeutic axis across different outcome modalities.
Feature Attribution: The stimulation-aligned axis was driven by interhemispheric PDC and high-beta power. In contrast, the therapeutic auxiliary dimensions were driven by low-beta burst dynamics, phase-amplitude coupling, and alpha-band power, linking them to parkinsonian motor pathophysiology rather than stimulation artifacts.
Superiority over Baselines: Under the same linear mixed-effects framework, unsupervised Principal Component Analysis (PCA) and canonical spectral band-power measures failed to recover the therapeutic structure or distinguish the anti-therapeutic 55 Hz condition.

Significance and Claims
The paper claims to offer a novel architectural template for discovering biomarkers in closed-loop settings where the recording and stimulation sites coincide. By "quarantining" the stimulation signature through supervised learning on a low-benefit condition, the method exposes a therapeutic structure in the orthogonal subspace without requiring outcome labels during the separation phase.

Decoupling Engagement from Benefit: The approach demonstrates that circuit engagement (detected by Dim 0) and therapeutic benefit (encoded in the auxiliary space) are distinct neural phenomena that can be separated mathematically.
Beyond Beta Power: The results suggest that therapeutic efficacy is not solely defined by beta suppression. The auxiliary metric captured nuances, such as the anti-therapeutic nature of low-frequency stimulation, which standard spectral summaries missed.
Generalizability: The frozen representation generalized across stimulation amplitudes, frequencies, patients, and even to a different outcome modality (kinematics vs. clinical rating), suggesting the learned geometry captures fundamental therapeutic principles rather than dataset-specific artifacts.
Future Application: The authors posit that this auxiliary axis could serve as a continuous feedback signal for adaptive DBS, allowing controllers to optimize the full stimulation parameter space (frequency and amplitude) rather than just amplitude based on beta suppression. This could prevent systems from reinforcing anti-therapeutic settings that merely engage the circuit.

The authors remain modest regarding the scope, noting that the findings are based on single-center data with small validation cohorts and that longitudinal stability remains untested. However, they assert that the method provides a robust proof-of-concept for separating intervention presence from intervention effect in neural recordings.