Hospitals cannot freely share scalp EEG. Recordings are protected health information, and they carry individual "brainprints" that let a classifier re-identify the person. The usual answer is to train a generative model on the private recordings and release synthetic signals instead. For multi-electrode EEG, though, the most useful clinical information is often not the raw waveform. It is the pattern of lagged synchrony between electrodes. The weighted phase lag index (wPLI) measures that pattern: it scores how consistently one electrode's rhythm leads or trails another's, and it ignores the instantaneous, zero-lag coupling that comes from electrical signal spreading through the skull (volume conduction). We propose to spend the formal privacy budget only on a small, bounded summary of each subject's wPLI network. Waveforms, if a user needs them, would then come from a public biophysical simulator driven by the privatized networks. Because a differentially private output stays private under any later processing, the waveforms inherit the guarantee at no extra cost.
The design follows from a tension the literature leaves open. Differential privacy is usually applied by clipping and adding noise to training gradients (DP-SGD). Applied to a waveform generator, that noise smears exactly the fine phase relationships that lag-based connectivity depends on. The alternative puts the noise where the quantity of interest is naturally bounded and low-dimensional. The core claim is that, at a matched subject-level privacy guarantee, connectome-first release preserves wPLI network structure and connectivity-based seizure prediction better than DP-trained waveform generators. That includes waveform generators given an explicit wPLI loss. The note also commits to reporting honestly how large a cohort this requires.
What exists and what is missing
Multi-channel neural generative models all justify themselves by clinical privacy, yet none implement formal differential privacy. Generating realistic neurophysiological time series with denoising diffusion probabilistic models trains diffusion models on EEG, ECoG and LFP and checks phase–amplitude coupling, but it never checks lag-based sensor networks and offers no (ε, δ) bound. EEG-DIF stacks 16 electrodes into an image for a diffusion model. That puts electrodes next to each other on the image grid regardless of where they sit on the scalp, and it has no privacy mechanism. Synthetic Sleep EEG Signal Generation using Latent Diffusion Models and EEG-GAN are single-channel, so they have no cross-electrode structure to preserve. These papers motivate the gap. Our method borrows from none of them except as a baseline.
Formal privacy for physiological time series does exist, but only for one channel. GlucoSynth splits its privacy budget between a teacher-ensemble mechanism on discrete motifs and DP-SGD on a latent model. It reports that plain DP time-series generators produce over-smoothed traces. In tabular clinical data, Synthetic data for privacy-preserving clinical risk prediction finds that DPGAN collapses and invents modes while PATE-GAN and ADSGAN hold up. These results constrain our design: they argue against injecting DP noise into a high-dimensional waveform model. We take one idea from GlucoSynth, which is to privatize a structured, low-dimensional component, and push it further by privatizing only that component.
On the connectivity side, the Kuramoto linear-mixing study shows that coherence and cross-correlation report full synchrony at zero true coupling once signals overlap, while PLI recovers true coupling transitions and clustering. Haufe et al. show that Granger-style directed measures are fooled by volume conduction and only the imaginary-coherence-based phase-slope index is not. These papers supply the rationale for privatizing a lag-based statistic, and they supply the simulators we reuse. An Explainable Statistical Method for Seizure Prediction shows that network features, not raw EEG, drive clinical prediction. That is the utility argument for making the connectome the released object. Its networks come from covariance, which is exposed to zero-lag leakage, so we treat it as the utility ceiling on real data rather than as the representation we adopt.
Core Claim
For one frequency band and electrode pair $(i,j)$, a subject's wPLI is $w_{ij} = |\langle \Im X_{ij} \rangle| / \langle |\Im X_{ij}| \rangle$. Here $X_{ij}$ is the cross-spectrum of the two electrodes on one epoch, $\Im$ takes its imaginary part, and $\langle\cdot\rangle$ averages over that subject's epochs. Every entry lies in $[0,1]$, so no gradient clipping or tuning is needed to bound one person's influence.
We project each subject's vector of edge values onto $k$ directions fixed in advance from simulated or public data. The projection matrix is orthonormal, so projection can only shrink norms. Each subject's projected vector therefore has $\ell_2$ norm at most $\sqrt{E}$, where $E$ is the number of edges across bands. For a mean over $n$ subjects, replacing one subject moves the mean by at most
$$\Delta_2 \le \frac{2\sqrt{E}}{n}.$$
The Gaussian mechanism adds noise of scale $\sigma = \Delta_2 \sqrt{2\ln(1.25/\delta)}/\varepsilon$, where $\varepsilon$ and $\delta$ are the usual privacy parameters. This worst-case bound does not improve by projecting to $k$ dimensions. The per-coordinate noise stays the same, but there are only $k$ noisy coordinates instead of $E$. A tighter bound in $k$ is available by clipping each subject's projected vector to norm $C$ before averaging. Its cost is some clipping bias, which we measure.
The arithmetic is sobering and we state it up front. With 16 electrodes (120 edges per band), four bands, 23 CHB-MIT subjects, $\varepsilon=1$ and $\delta=10^{-5}$, the per-coordinate noise scale is several times larger than typical wPLI values, even with clipping. Subject-level private release of connectomes is therefore feasible only with few released coordinates, pooled cohorts, or larger $\varepsilon$. Finding the smallest cohort at which it works is itself a result. A waveform-level DP-SGD model faces the same small $n$ with far more parameters, which is why we expect connectome-first release to win.
Research Program
Stage one is the private connectome generator. At each site, per-subject wPLI matrices are computed in delta, theta, alpha and beta bands under a fixed montage and projected to $k$ coordinates. Within each clinical class, such as preictal versus interictal windows, we release the privatized mean and covariance of the projected coordinates, or a small privatized Gaussian mixture. Sampling from this release yields synthetic connectomes.
The generator is deliberately simple. The contribution is where the noise is placed, not model capacity. The projection basis is learned only from Kuramoto and Jansen-Rit simulations (Perturbing a Neural Network to Infer Effective Connectivity) or from a disjoint public cohort, so it consumes no privacy budget.
Stage two is an optional public waveform decoder. An amortized inverse network is trained purely on simulations. It maps a target wPLI matrix and target band powers to the coupling and delay parameters of a Kuramoto or Jansen-Rit source network. The simulated source signals are then mixed onto the electrodes through a fixed template lead field, a matrix describing how each brain source projects to each scalp electrode, following the mixing setup of the Kuramoto study. Because volume conduction is added explicitly, released waveforms show realistic zero-lag inflation in coherence while carrying the privatized lagged structure in wPLI. The decoder never touches private data, so it adds no privacy cost.
The analysis crosses three factors. The first is the privacy unit: subject level, and recording-hour level as a weaker, clearly labelled variant. The second is ε, from 0.5 to 8. The third is cohort size, using real cohorts plus simulated populations of 20 to 2000 subjects, so we can read off how fidelity scales with n.
Baselines
The naive reference points bracket the problem. The first is releasing nothing: a downstream classifier gets only public simulated data. The second is a non-private oracle trained and analyzed on the real recordings. Every method is scored on how much of the gap between these two it closes.
The strongest existing method is the multi-scale diffusion model of Vetter et al., retrained with DP-SGD at the same (ε, δ) and the same privacy unit. It is the most capable multi-channel neural generator in the file. On the non-private side it reports 84% accuracy for train-on-synthetic, test-on-real brain-state classification. No reported result was found on Lacuna for any DP-trained multi-channel EEG generator, so we must establish its private performance ourselves. It gets the same data, a matched GPU-hour budget, and the same hyperparameter search budget, with privacy accounting that includes the search.
The obvious-idea baseline adds a wPLI-matching loss to that DP-SGD diffusion model. It tests whether telling a waveform model about phase lag is enough, or whether noise placement is what matters.
The key-ingredient ablation runs our pipeline with magnitude-squared coherence instead of wPLI as the privatized statistic. This isolates the value of a lag-based, volume-conduction-resistant summary.
For utility on real data, the reference is the precision-matrix network predictor of Sun et al., which reports AUC 0.940, 93.0% sensitivity and 11.1% false discovery rate on CHB-MIT. For waveform generation on Siena, the reference is EEG-DIF, which reports downstream seizure-prediction AUC 0.89 without privacy.
Beating the strongest competitor means two things at every tested ε ≥ 1 with subject-level privacy. The synthetic wPLI networks must match real clustering coefficient and betweenness centrality significantly better than the DP diffusion model does. And a seizure predictor trained on synthetic connectomes and tested on real held-out patients must reach a significantly higher AUC than the same predictor trained on DP-diffusion output.
Why not just…?
Why not just release the privatized wPLI matrices, or a DP Gaussian over them, and skip the waveforms? This is a real threat, and it is stage one of our own pipeline. If downstream users only ever need connectomes, stage two adds nothing. We will report stage one alone as a first-class result. The waveform decoder is justified only if users need raw signals, for example to re-bandpass, change the reference, or compute measures such as the phase-slope index that are not in the released summary. We test this by computing unreleased measures on decoder output and comparing them with real data. If those measures are no better than simulated noise, we narrow the claim to connectome release.
Why not just train a waveform diffusion model with DP-SGD and a wPLI loss? The loss tells the model what to preserve, but the gradient noise is still spread over millions of parameters. GlucoSynth and the UK Biobank study both report utility collapse from exactly this. The head-to-head comparison at matched ε decides the question. If this baseline matches us, noise placement is not the operative ingredient, and we would say so.
Why not just de-identify the recordings or strip identity features, as ID-RemovalNet does? ID-RemovalNet reduces re-identification accuracy to 0.43% inside a closed decoder. It does not release data and gives no formal bound. We still use its framing for evaluation: a trained identity classifier and a membership-inference attack are run against every released artifact.
Why not federated learning? It trains one fixed model and does not support the exploratory network analysis and biomarker search that clinicians run on connectomes, as the UK Biobank synthetic-data study argues. It also sits beside our proposal rather than competing with it on the question asked here.
Proposed Experiments
The decisive checks run on CHB-MIT and Siena, pooled where montages allow, and on the BCI Competition IV-2a motor imagery data with 22 electrodes. Simulated Kuramoto and Jansen-Rit populations with known ground-truth coupling let us vary n far beyond the tiny real cohorts. We are aware that Siena has only 14 patients and IV-2a only 9, so subject-level results there serve as a stress test rather than a showcase.
Fidelity is measured in three ways: edge-wise error of wPLI against held-out real subjects, graph-invariant agreement, and, in simulation, recovery of true coupling. Utility is measured by seizure prediction AUC and sensitivity from connectome features under train-on-synthetic, test-on-real. Privacy is measured by the formal ε, plus empirical identity-classifier accuracy and membership-inference advantage.
The first decisive check asks where the curves of fidelity against ε for connectome-first release and DP diffusion cross, as a function of n. Our hypothesis predicts that connectome-first release dominates at small n and the gap closes as n grows.
The second checks the physics of the decoder output. Coherence should show strong zero-lag structure while wPLI does not, reproducing the real-data dissociation that the Kuramoto study predicts. The DP diffusion baseline, we expect, will show spurious lagged edges instead.
The third reports the minimum pooled cohort size at which subject-level ε = 1 release keeps seizure-prediction AUC within 0.05 of the real-data oracle. If no feasible cohort in the file reaches that threshold, that negative result is itself the main finding for anyone planning private EEG sharing.
References
- Generating realistic neurophysiological time series with denoising diffusion probabilistic models
- EEG-DIF: Early Warning of Epileptic Seizures through Generative Diffusion Model-based Multi-channel EEG Signals Forecasting
- Synthetic Sleep EEG Signal Generation using Latent Diffusion Models
- EEG-GAN: Generative adversarial networks for electroencephalograhic (EEG) brain signals
- GlucoSynth: Generating Differentially-Private Synthetic Glucose Traces
- Synthetic data for privacy-preserving clinical risk prediction
- Effect of linear mixing in EEG on synchronization and complex network measures studied using the Kuramoto model
- Identifying brain effective connectivity patterns from EEG: performance of Granger Causality, DTF, PDC and PSI on simulated data
- An Explainable Statistical Method for Seizure Prediction Using Brain Functional Connectivity from EEG
- Perturbing a Neural Network to Infer Effective Connectivity: Evidence from Synthetic EEG Data
- ID-RemovalNet: Identity Removal Network for EEG Privacy Protection with Enhancing Decoding Tasks