From a Locust Antenna to a Bio-Hybrid "Nose": Leakage-Resistant Odor Decoding with Grouped Evaluation, Time–Frequency Features, and Interpretable Dynamics

Bio-hybrid sensing—pairing biological sensing organs with machine learning—offers a promising alternative to conventional electronic noses, particularly where engineered sensors remain drift-prone, costly, or less sensitive than natural olfaction. Electroantennogram (EAG) recordings from insect antennae provide a simple and scalable interface to olfactory receptor activity, but evaluation in biosensing is highly vulnerable to information leakage (e.g., repeated-measure artifacts, session-specific signatures, and correlated replicates crossing train/test boundaries). We present a leakage-resistant generalization study for odor decoding from locust EAG signals inspired by Neta et al. (2023). The protocol enforces strict group-wise splitting by recording unit (“ID”), DEV/TEST separation with a one-shot locked TEST evaluation, and nested cross-validation restricted to DEV. We quantify uncertainty via bootstrap resampling at the ID level and validate time-resolved decoding with permutation tests using max-stat correction across time. For interpretability, we derive time–frequency features using sliding-window Welch power spectral density (PSD; 1–20 Hz) and dynamic ΔPSD representations, enabling ablations over frequency bands and time segments and weight backprojection into time–frequency importance maps. Across datasets comprising eight single odors and two mixture settings (two components plus their binary mixture), tree-based waveform baselines confirm decodability under strict grouped validation on DEV, while a locked linear pipeline evaluated once on held-out IDs yields balanced accuracies consistent with honest generalization (≈0.67 on the 8-odor holdout; ≈0.68–0.72 on mixture tasks with bootstrap-by-ID uncertainty). These results provide a reproducible blueprint combining leakage-resistant evaluation, interpretable dynamics, and robust statistics to support credible progress from lab demonstrations toward deployable bio-hybrid odor sensing systems.