The Science Behind SynapCore
How SynapCore turns a raw EEG recording into a decision-support report, and how we test it. Every step below describes what the code does today.
From Raw Signal to Clinical Insight
Five stages turn a multichannel EEG recording into a structured decision-support report.
Raw EEG Acquisition
EEG recordings uploaded as EDF files. The pipeline uses the 19 standard channels of the 10-20 system and needs at least 15 of them to run.
Preprocessing
Resampling to 256 Hz, 1-45 Hz band-pass and 50 Hz notch filtering, 2-second epochs, and automatic epoch cleaning with AutoReject, which repairs or drops noisy segments.
Feature Extraction
Absolute and relative power in the delta, theta, alpha, beta and gamma bands, theta/beta and alpha/theta ratios, coherence between channel pairs, and complexity measures (entropy, Hjorth parameters, DFA).
ML Classification
LightGBM gradient-boosted classifier, hyperparameters tuned with Optuna, probabilities calibrated with a sigmoid calibrator. Evaluated with leave-one-subject-out (LOSO) cross-validation.
Clinical Report
A calibrated probability and the features that drove it (SHAP), written up as a structured report. Designed for clinician review, not autonomous diagnosis.
ML Architecture
SynapCore uses LightGBM, a gradient-boosted decision tree model, chosen because it works well on tabular EEG features with small cohorts and its predictions can be explained feature by feature.
LightGBM Classifier
Gradient-boosted trees trained on the extracted EEG features. Each prediction comes with SHAP values that show which features pushed the probability up or down.
LOSO Cross-Validation
Leave-one-subject-out validation keeps every recording of the test subject out of training. It gives a more conservative and realistic estimate of performance on a new person than k-fold splits.
Clinical Relevance
LOSO mimics the real situation, where the model sees a patient it has never seen before. That is why we report LOSO results rather than the higher k-fold scores common in EEG-ML papers.
Performance Metrics
Neuraxis evaluated on public research datasets with LOSO cross-validation. All metrics are preliminary research results.
| Metric | Neuraxis-Child v1.0Ages 7-12 | Neuraxis-Adult v1.0Ages 18-35 · under revalidation |
|---|---|---|
| AUC | 0.650 | Not reported |
| Balanced Accuracy | 0.595 | - |
Internal LOSO results on the IEEE cohort. On an independent external cohort (HBN-EEG, N=321, ages 5 to 17) Neuraxis-Child scored AUC 0.519, which is chance level. All reported metrics are preliminary research results. Research Use Only (RUO).
Research Datasets
Developed and tested on publicly available EEG datasets, so the work can be checked and reproduced.
IEEE ADHD Dataset
Publicly available paediatric EEG dataset used for ADHD classification research, with clinical ADHD diagnoses. Primary training and validation set for Neuraxis-Child.
TDBRAIN
Lifespan EEG dataset (ages 5 to 89) from the Brainclinics Foundation, Nijmegen, The Netherlands (van Dijk et al., 2022). We selected a subset of subjects within the adult ADHD age range. Used in Neuraxis-Adult validation work with LOSO cross-validation; the adult module is under revalidation.
HBN-EEG
Healthy Brain Network resting-state EEG (105 ADHD, 216 control), used only as an independent external test of Neuraxis-Child. No HBN subject was used in training. The model scored at chance level on it (AUC 0.519).
Regulatory Status
SynapCore is a research tool. It has not been assessed by any regulator and must not be used to diagnose or treat patients.
Current status
NowResearch Use Only. Not UKCA marked, CE marked or FDA cleared.
Before clinical use
RequiredSoftware that supports diagnosis is regulated as a medical device. SynapCore would need UKCA marking to be used in the UK, and CE marking or FDA clearance for the EU or US.
How we build it
NowVersioned code, reviewed changes and documented, audited methods. A formal IEC 62304 process has not started yet.
Next step
PlannedValidation on clinical data with a research partner, under ethical approval, before any regulatory submission.
Working in EEG research or clinical neuroscience?
We are looking for research and clinical partners to help validate SynapCore.