Neurostimulation & Language Database — Elements of the Neurobiology of Language
Language Elements is a living systematic review database of neurostimulation studies examining the causal role of brain regions in language processing. It compiles findings from transcranial magnetic stimulation (TMS), transcranial electrical stimulation (tES), and direct electrical stimulation (DES) studies — the only techniques capable of establishing causal brain–language relationships rather than correlational ones.
The database is built around elementalism: a theoretical framework that characterises brain regions by the minimal computational operations they causally support, inferred bottom-up across heterogeneous tasks (following Genon et al., 2018). This high-specificity approach distinguishes Language Elements from existing resources, which typically catalogue findings at the level of broad linguistic domains.
The database is designed to serve both basic research — supporting experimental planning and meta-analytic synthesis — and intraoperative clinical decision-making, providing evidence-based task recommendations for awake craniotomy language mapping.
The systematic review (PROSPERO: CRD42024602006) searched PubMed, Scopus, Embase, and PsychInfo from October 2024 to January 2025, returning 12,763 records. After deduplication and screening, the current database includes:
The database is live and updated as screening and extraction continue.
A paper describing the database and the elementalism framework is currently in submission. In the meantime, please cite the PROSPERO registration:
Language Elements is an international collaboration between research teams in the UK and Germany.
This work was supported by the British Council Going Global Partnerships Springboard Programme (UK–Germany).
For queries about the database, the systematic review, or potential collaborations, please contact T. R. Williamson at t.williamson@uwe.ac.uk.
Neuroscientist Mode searches the database directly. Every result is drawn verbatim from the systematic review dataset — no inference, no generation. Free-text search and filters (stimulation type, linguistic area, hemisphere, inhibition/facilitation) operate on the raw data. The brain visualiser plots MNI coordinates extracted directly from the included papers.
Type a brain region name to begin. The tool runs in two stages: first characterising the functional profile of the region, then generating intraoperative task recommendations based on that characterisation.
When you search a region, the tool identifies all neurostimulation studies in the database targeting that region and extracts the linguistic processes they implicate. These processes are then grouped into operation-function labels using one of three layers, each with a defined fallback:
Processes are matched against a curated controlled vocabulary of operation-function labels. If coverage is sufficient, results are grouped instantly using this vocabulary alone — no network request is made. This layer is fully deterministic and reproducible across sessions.
When controlled vocabulary coverage is insufficient, an AI model (Claude, Anthropic) infers operation-function labels from the identified processes, following the bottom-up characterisation framework of Genon et al. (2018). The model receives only the list of processes and the region name — it does not access individual paper content. Results may vary slightly between sessions due to the probabilistic nature of language models.
If neither layer 1 nor layer 2 is available — because the AI call fails, the controlled vocabulary has insufficient coverage, or there is no internet connection — the tool displays all identified processes in a flat ranked list ordered by study count. All data remains drawn directly from the systematic review.
Once a region has been characterised, clicking Generate Recommendations runs a three-step process:
All tasks used in neurostimulation studies of this region are retrieved from the database and organised by the processes they target. Tasks are deduplicated and prepared for the AI, which receives the full task evidence for this region alongside the functional characterisation from the previous stage.
An AI model selects and ranks tasks from the database evidence, applying clinical constraints from two peer-reviewed guardrail documents loaded at runtime:
The AI is instructed to recommend tasks from database evidence only, flag data gaps explicitly, label every recommendation by evidence type and DES compatibility confidence, and never extrapolate beyond the evidence base. All recommendations cite their source papers.
If the AI call fails — due to a network error, a timeout, or a service interruption — you will see: "Failed to generate recommendations. Please check your API key and try again." The tool does not produce unsupported output. No recommendations are shown unless they are grounded in the database evidence and the clinical guardrail documents. In this case, the functional characterisation from the previous stage remains visible and can still be used to inform clinical judgement.
They are not validated clinical risk scores and should not replace professional surgical judgement. The raw confidence score ranges from 1 to 7 points (up to 5 for study count, +1 if any study used DES, +1 if total N > 50); this is scaled to a 1–5 dot display. A score of ●●●●● therefore represents strong relative evidence within this dataset — it does not indicate perfect or complete evidence.