Research
Temporal interference stimulation with concurrent intracranial recording
Deep structures in the brain are hard to reach without surgery. Current applied at the scalp spreads out on the way in, so the tissue you were aiming for receives the least of it and everything above receives the most. That is the constraint every noninvasive stimulation method runs into.
Temporal interference takes a different route. Two fields are applied at the scalp, each oscillating too quickly for neurons to follow. Where the two fields overlap they beat against one another, and it is that slow beat — not either field alone — that the tissue responds to. Place the overlap correctly and the effect can land at depth while the surface is left comparatively untouched.
Whether that holds in a living human brain is a separate question from whether it holds in a model. This project puts it to patients who already have electrodes implanted for clinical seizure monitoring. While those electrodes are in place, they record what the stimulation is doing at the structures it was aimed at — measured rather than inferred. The first human study to come out of that setup, published in Brain Stimulation, found that stimulation delivered this way suppresses epileptic biomarkers in the hippocampus.
Methods and technical detail
Design
Within-subject stimulation protocol run during clinical intracranial monitoring, with concurrent recording from implanted electrodes. Stimulation and sham blocks are interleaved with cognitive task blocks so that electrophysiological and behavioural effects are captured in the same session.
Participants
Patients undergoing intracranial monitoring for seizure localisation, enrolled on a research protocol alongside their clinical care. Electrode placement is determined entirely by clinical need; the research protocol adds no implants and drives no surgical decision.
Acquisition
Concurrent scalp stimulation and intracranial recording. Recordings include iEEG, ECoG, and SEEG depending on the implant a given patient has, at clinical sampling rates, which puts individual sessions into the multi-gigabyte range and makes streaming and chunked reads a requirement rather than an optimisation.
Preprocessing
Bandpass filtering, notch filtering at line frequency and its harmonics, and bipolar re-referencing against the adjacent contact on each electrode shaft. Bipolar referencing matters more than usual here: it suppresses the shared stimulation artefact, which is common-mode across nearby contacts, while preserving the local signal that the analysis is actually about. Channels are screened for clinical artefact and for contacts in the seizure onset zone before any group analysis.
Analysis
Power spectral density estimation across canonical bands, with the beat frequency and its neighbourhood treated as the band of interest. Comparisons are within-subject between stimulation and sham, contact by contact, with distance from the intended target as the organising variable — the prediction being that the effect grows toward the target rather than falling off from the scalp.
Stack
Python with MNE-Python for the primary pipeline; MATLAB with EEGLAB for cross-checking preprocessing decisions against the lab’s existing workflows. R and SPSS for the statistical modelling. Histological validation runs alongside the human work — mouse perfusion and brain slice immunostaining to visualise stimulation effects in hippocampal tissue directly. The import layer is the part that took real engineering: clinical recording systems write large proprietary files, and getting them into an analysable form reproducibly, without loading a whole session into memory, is most of the work before any analysis starts.