Interpretable machine learning of micro-ERG reveals muted retinal gain tracking in 5xFAD mice

Tue, 01 Sep 2026 00:00:00 +0000

Funding

Becas Chile de Postdoctorado en el Extranjero; ANID Exploración 13220082; Fondecyt Regular 1241202.

Interpretable machine learning of micro-ERG reveals muted retinal gain tracking in 5xFAD mice

Mon, 31 Aug 2026 09:00:00 -0400

Background. Inexpensive, scalable biomarkers of early Alzheimer’s disease (AD) remain an unmet need, and the retina, as a developmental enlargement of the central nervous system, may offer such a window. The 5xFAD mouse exhibits electrophysiologically detectable retinal alterations. We ask whether multi-electrode micro-electroretinogram (μERG) signals can discriminate wild-type (WT) from 5xFAD retinas, and whether deep-learning interpretability may guide the discovery of mechanistic features.

Methods. We recorded μERG from 23 WT and 23 5xFAD, across young and adult cohorts, using a 16x16 multi-electrode array under a 35-second chirp (flash, frequency, and amplitude sweeps) and a 10-second natural-image video. Under subject-disjoint 5-fold cross-validation with age-blind inputs, we compared three strategies: 1) 1-D convolutional neural networks (CNNs), 2) classical machine learning on hand-crafted (HC) features, and 3) multiscale-entropy complexity features.

Results. For chirp, HC with logistic regression reached a pooled area under the curve (AUC) of 0.735, complexity features 0.729, and CNN 0.590; for natural images, the values were 0.782, 0.736, and 0.752. Further, interpretability analyses, i.e., Grad-CAM, integrated gradients, virtual band blockade, and Bayesian input optimization, localized the CNN’s cue to the amplitude-sweep fundamental-frequency gain trajectory, a pattern we term “muted gain tracking” in 5xFAD; encoding it as four additional features raised chirp AUC to 0.834, thereby surpassing the network.

Conclusions. Taken together, μERG signals carry a robust 5xFAD signature recoverable by machine learning, with compact, biologically-grounded features matching end-to-end deep learning; nevertheless, CNN interpretability translates modest performance into mechanistic markers, namely, retinal gain tracking, with promising translational potential as AD biomarkers.

Financiamiento/Funding. Becas Chile de Postdoctorado en el Extranjero; ANID Exploración 13220082; Fondecyt Regular 1241202.

Joseph West | NeuroEng@Usach

Interpretable machine learning of micro-ERG reveals muted retinal gain tracking in 5xFAD mice

Funding

Interpretable machine learning of micro-ERG reveals muted retinal gain tracking in 5xFAD mice