Interpretable Machine Learning Identifies Electrophysiological Signatures of SYNGAP1 Genotype and Conditioned Freezing

We apply interpretable machine learning to multi-region neural recordings from a rat model of SYNGAP1 haploinsufficiency to investigate whether genotype-associated electrophysiological patterns can support classification of previously unseen animals. EEG/LFP activity from the medial prefrontal cortex, amygdala, and olfactory bulb was represented using spectral, nonlinear complexity, and functional connectivity features. Supervised classifiers were compared using nested paired leave-two-animals-out cross-validation, with one wild-type and one Syngap+/Δ-GAP animal held out in each outer fold. XGBoost achieved balanced accuracies of 0.857 during habituation and 0.750 during fear recall. Genotype remained decodable when recall data were analysed separately during freezing and non-freezing periods, indicating that performance was not driven solely by behavioural differences. SHAP analysis identified recurring gamma-related features, including amygdala gamma power and amygdala–mPFC gamma coherence, across independently trained models. These preliminary results illustrate how animal-level validation, behavioural-state stratification, and model attribution can be combined to study genotype-associated neural signatures in small-cohort, repeated-measures datasets. Ongoing analyses are evaluating the robustness and biological relevance of these candidate signatures.