NeuroSim: a clinically informed multi-disorder EEG simulator with personalized pattern injection and AI-driven neurological modeling

Avatar
Poster
Voice is AI-generated
Connected to paperThis paper is a preprint and has not been certified by peer review

NeuroSim: a clinically informed multi-disorder EEG simulator with personalized pattern injection and AI-driven neurological modeling

Authors

Bunterngchit, C.; Aslam, S.; Sattar, A.; Rasool, A.; Wang, Y.

Abstract

Deep learning for electroencephalography (EEG) based neurodiagnostics is constrained by the limited availability of labeled clinical recordings. Datasets covering Alzheimers disease (AD), Parkinsons disease (PD), schizophrenia (SCZ), epilepsy (EPI), and frontotemporal dementia (FTD) are scarce, demographically narrow, and heterogeneous across acquisition sites, while privacy requirements restrict data sharing. Existing EEG simulators typically address a single disorder or trade biological interpretability for statistical fidelity. This study presents NeuroSim, a simulation framework that generates synthetic multichannel EEG with user-controlled injection of disease-specific patterns across five disorders and multiple severity levels. A physiologically motivated baseline generator produces band-limited rhythms, event-related potentials, artifacts, and hierarchical subject-level and trial-level variability. Disease patterns are introduced through region-specific and severity-specific spectral modulation that encodes established biomarkers, including posterior alpha suppression in AD, motor beta enhancement in PD, and spike-wave discharges in EPI. An interactive interface exposes montage, sampling rate, duration, artifacts, stimuli, affected regions, and severity, and exports signals with reproducible metadata. The simulated signals reproduced the expected biomarker directionality under spectral, topographic, and multivariate statistical analysis. A convolutional-attention classifier trained on simulated data reached 95.3 to 96.8 percent hold-out accuracy, and post hoc explainability analyses produced class-separable embeddings with attributions concentrated on pathophysiologically expected electrodes. Application of the analysis pipeline to recordings from the Temple University Hospital EEG Corpus confirmed compatibility with heterogeneous clinical data. In subject-level augmentation experiments on real clinical EEG, supplementing real training data with NeuroSim samples increased classification accuracy from 91.6 to 94.5 percent for AD versus healthy controls and from 90.9 to 94.0 percent for PD versus healthy controls, exceeding the improvement obtained with conventional signal-level augmentation. NeuroSim is released as open-source software to support augmentation, benchmarking, and explainable artificial intelligence in neurological diagnostics.

Follow Us on

0 comments

Add comment