People with Parkinson's disease sometimes "freeze" mid-walk: their feet seem glued to the floor even though they intend to keep moving. To a motion sensor, a freeze can look exactly like a deliberate stop. That ambiguity matters for assistive devices, such as cueing systems or exoskeletons, that must help during a freeze but stay out of the way when the person simply wants to stand. Scalp electroencephalography (EEG, electrical brain activity recorded from the scalp) seemed to offer a way through. In a 2-second window before the event, freezes and voluntary stops show different oscillatory power, with a rise in central beta (13–30 Hz) before freezes and broad power increases before stops (Identification of EEG Dynamics During Freezing of Gait and Voluntary Stopping). Deep networks separate the two classes at 87–88% accuracy across held-out patients (Predicting the Onset of Freezing of Gait Using EEG Dynamics).
This proposal argues that the evidence is weaker than it looks, and it sets out a way to find out how much of the signature is real. In the source protocol, every voluntary stop follows a spoken "stop" command, while freezes cluster at turns and doorways. A decoder could therefore separate "just heard a word" from "turning the head and body" without reading anything about volition or motor breakdown. The class balance adds a second problem. With 178 freeze epochs and 55 stop epochs, always guessing "freeze" already scores about 76%, so the reported 87–88% sits only about 11 points above the trivial reference. We propose an interventional audit with three parts: a recording protocol that pulls apart cue, context, and stopping; targeted edits to the signal that remove one ingredient at a time and measure how much the decoder's output depends on each; and a decoder trained to ignore the confounds and tested on situations the confounded decoder cannot have learned. The deliverable is an honest decomposition of the freeze-versus-stop separation into four parts: cue, movement context, background spectral shift, and oscillatory power within specific frequency bands. Any later brain-stimulation trial would need this decomposition to choose its target.
Core Claim
Our hypothesis, which can be proven false, is this. Once auditory cueing and turning context are matched, the gap between freeze and voluntary-stop transitions in held-out patients will shrink substantially. Most of what survives will come from two sources. The first is the aperiodic background: the smooth 1/f tilt of the spectrum, as opposed to oscillatory peaks. The second is low-frequency midfrontal activity. Central beta, the feature usually emphasized, is expected to contribute less.
The prediction draws on a tension already present in the file. The resting-state and timing literature points to midfrontal delta (1–4 Hz) as the stronger marker of freezing vulnerability: delta reaches an AUC of 0.93 against 0.81 for theta in Interval timing and midfrontal delta oscillations are impaired in PD patients with freezing of gait. Midfrontal beta reaches only 0.68 AUC at rest in EEG-Based Classification of PD With Freezing of Gait Using Midfrontal Beta Oscillations. Ordinary power measures also mix background tilt with oscillatory peaks (A Computational Framework for EEG Causal Oscillatory Connectivity).
If the hypothesis fails, and central beta keeps its discriminative weight under matched conditions and survives our edits, that is an equally useful result. It would justify beta as a target for later stimulation.
We are explicit about which level of causal evidence this work reaches. The edits we apply are interventions on the decoder and on the recorded data, not on the brain. In the terms of Causality is Key for Interpretability Claims to Generalise, which found that most interpretability claims assert more causal strength than their evidence supports, our edits establish what the model uses (an interventional, "rung 2", claim about the decoder). The protocol design gives a quasi-experimental test of what the brain signal tracks. Neither says that beta or delta drives the freeze. We will not claim that it does.
Research Program
The recording protocol. We reuse the Timed Up-and-Go paradigm, a standard clinical test in which the person stands, walks, turns, and returns, with 32-channel mobile EEG recorded in patients off their dopaminergic medication. This is the setup of the John and Cao studies above. We add four trial types that break the confounds.
The first is self-initiated stops at a floor marker with no spoken cue, which removes the auditory response from voluntary stopping. The second is spoken "walk" or neutral-word catch cues delivered during steady walking, which present an auditory command without any stop. The third is "stop" cues delivered during turns, so that cued stops and freezes share the same movement context. The fourth is turning segments that do not freeze, which serve as context-matched controls for freeze transitions.
We also record eye and neck-muscle reference channels and a head-mounted inertial sensor (an accelerometer and gyroscope that measure motion). This answers the concern raised in Assessing the Capabilities of Large Brainwave Foundation Models that EEG decoders exploit eye artifacts. A subset of patients repeats the protocol on medication. That addresses the off-versus-on disconnect between the dynamic studies and the resting-state work, and it tests whether a surviving signature persists under real-world dopaminergic conditions.
We will request the original dataset from its authors so we can reproduce their baseline. The file does not say whether it is public.
Signal decomposition. Following A Computational Framework for EEG Causal Oscillatory Connectivity, we first apply a surface Laplacian, a spatial filter that reduces blurring from electrical spread through tissue, skull, and scalp. We then use spectral parameterization ("specparam") to split each window's spectrum into an aperiodic component (offset and slope) and periodic peaks per band. That paper showed that this separation, combined with directed causal discovery, detected treatment effects that ordinary coherence missed. We reuse the decomposition as an editing tool rather than only as a descriptive one.
Interventions on the decoder. For a trained decoder $f$ that outputs the log-odds of "freeze" for a pre-transition window $x$, we define the effect of ingredient $k$ as
$$\Delta_k = \mathbb{E}_x\big[f(x) - f(T_k(x))\big],$$
where $T_k$ is an edit that removes only ingredient $k$ while keeping the rest of the window intact. The edits are as follows. To remove a band's periodic peak, we replace it with a surrogate whose phases are randomized but whose aperiodic fit is unchanged. To flatten the aperiodic slope toward the pooled mean, we leave the peaks in place and reconstruct the signal. To remove artifacts, we regress out the eye, neck-muscle, and head-motion channels. To remove the cue, we subtract a template of the auditory evoked response estimated from the catch-cue trials.
The expectation is taken over held-out patients, and each $\Delta_k$ is reported relative to the size of the total logit separation between classes. This mirrors the perturbation analyses of Hierarchical internal representation of spectral features in deep convolutional networks trained for EEG decoding, which showed that convolutional networks rebuild frequency-band structure inside their layers. Here the perturbations are aimed at specific confounds rather than at general interpretability.
A confound-invariant decoder. We train the Deep ConvNet used by John et al., adding subject-conditioned dynamic convolutions. These are layers whose filter weights are mixed differently for each subject, following A Causal Viewpoint on Motor-Imagery Brainwave Decoding, where they raised cross-subject accuracy from 67.1% to 71.3%. We also add a penalty that makes the decoder's representation uninformative about cue presence and turning state. The penalty is fit on the catch and turn-control trials, where those variables vary independently of the stop/freeze label.
Baselines
The naive reference is majority-class prediction. On the original epoch counts it scores roughly 76% accuracy for freeze versus stop. This figure is derived from the 178 freeze and 55 stop epochs reported in Predicting the Onset of Freezing of Gait Using EEG Dynamics. Because of that baseline, all our comparisons use balanced accuracy and the Matthews correlation coefficient (a correlation-based score that stays fair under class imbalance), not raw accuracy.
The strongest existing method is the Deep ConvNet of John et al., trained on 2-second, 1–30 Hz windows. The paper reports 87–88% two-class accuracy and 70–74% three-class accuracy under leave-one-subject-out cross-validation (holding out each patient in turn). We first reproduce this number on the original data, then retrain the model on our protocol with the same windows, preprocessing, and hyperparameter search budget as every other model.
A third baseline is a context-only decoder. It sees no brain-frequency EEG content, only eye, neck-muscle, and head-motion channels plus indicators for "cue in the last 2 s" and "turning". This is the skeptic's model. If it approaches the Deep ConvNet, then the EEG result is largely explained by artifacts and context.
A fourth baseline is an ankle and waist inertial-sensor decoder in the style of FM-FoG, which reports a 98.5% F1 score for detecting freezes in unseen patients. That figure is for detecting a freeze, not for predicting freeze versus stop before onset. We know of no reported result for that pre-onset task.
The fifth baseline is a logistic regression on specparam features: periodic band power plus aperiodic slope and offset at Fz and Cz.
The key ablation compares the invariant decoder with and without the confound penalty, and with periodic-only versus aperiodic-only inputs. This isolates the proposal's main ingredient.
"Beating" the strongest baseline means one thing here. On the matched test set, where cued stops occur during turns and uncued stops occur during straight walking, the invariant decoder achieves higher balanced accuracy than the Deep ConvNet across held-out patients, with a paired test across patients. It must also exceed the context-only decoder by a margin that does not overlap across bootstrap resamples.
Why not just…?
Why not just use inertial sensors, since they detect freezes nearly perfectly? Inertial-sensor detectors work after movement has already failed, and a still body looks the same whether the stop was chosen or not. They may still be a real threat in the pre-onset window: approach slowing and turning already precede many freezes. If the inertial-sensor baseline matches EEG on our matched test set, we will report that EEG adds nothing for this decision. That outcome would be informative for device design.
Why not just run the causal oscillatory connectivity pipeline on the existing data? Directed connectivity between scalp regions can be estimated perfectly well and still reflect the auditory command, because the confound sits in the task design, not in the analysis. The method of A Computational Framework for EEG Causal Oscillatory Connectivity was also validated only at rest. We use its decomposition step and add a protocol that breaks the cue-label link. Its directed-graph step is a natural follow-up once a confound-free contrast exists.
Why not just clean the artifacts with independent component analysis or artifact subspace reconstruction? These methods remove eye and muscle signals, as in the real-time pipeline of A 2-Block Architecture for Real-Time EEG Gait Decoding. They do not remove genuine brain responses to a spoken word or to turning, and those responses are the central confound here. The catch-cue trials measure this directly. If a cleaned decoder labels "walk" cues as stops at a high rate, cleaning was not enough.
Why not go straight to stimulating central beta and see whether freezing changes? Phase-specific stimulation can establish causation, as theta dual-brain stimulation did for the sense of joint agency (Theta Dual-Brain Stimulation of rTPJ Shapes Joint Agency). The trouble is that the file leaves open whether beta or delta is the right target. A trial run on a confounded marker would waste patient effort. This study is meant to choose that target.
Decisive Checks
The first decisive test is transfer. A decoder trained on the original cued data is applied to uncued self-initiated stops and to cued stops during turns. If its balanced accuracy collapses toward chance on these, while the invariant decoder holds, the original signature was largely cue and context.
The second test is the catch-cue false-alarm rate: the fraction of "walk" cues that the decoder labels as stop transitions.
The third test is the set of $\Delta_k$ values. They tell us whether the surviving separation rests on aperiodic slope, midfrontal delta, or central beta. We will also check whether this ranking holds on medication.
For deployment, we report latency on windows shorter than 1 second. The 2-second offline window is too slow for closed-loop help. The only real-time gait decoder in the file reached about 1.03 s end-to-end at a Matthews correlation coefficient of 0.435, in one healthy participant.
References
- Identification of EEG Dynamics During Freezing of Gait and Voluntary Stopping in Patients With Parkinson’s Disease
- Predicting the Onset of Freezing of Gait Using EEG Dynamics
- Interval timing and midfrontal delta oscillations are impaired in Parkinson’s disease patients with freezing of gait
- EEG-Based Classification of Parkinson’s Disease With Freezing of Gait Using Midfrontal Beta Oscillations
- A Computational Framework for EEG Causal Oscillatory Connectivity
- Causality is Key for Interpretability Claims to Generalise
- Assessing the Capabilities of Large Brainwave Foundation Models
- Hierarchical internal representation of spectral features in deep convolutional networks trained for EEG decoding
- A CAUSAL VIEWPOINT ON MOTOR-IMAGERY BRAINWAVE DECODING
- FM-FoG: A Real-Time Foundation Model-based Wearable System for Freezing-of-Gait Mitigation
- A 2-Block Architecture for Real-Time EEG Gait Decoding: A Pilot Study
- Theta Dual-Brain Stimulation of rTPJ Shapes Joint Agency