Spontaneous ERG-based random forest model discriminates control and disease cases in rodent models of obesity and type 1 diabetes and predicts disease evolution in the type 1 diabetes model.

Illustrative spontaneous ERGs and wavelet analysis in A, control versus high-fat diet-fed mice (n = 75 and n = 75, respectively), C, lean versus spontaneously obese Neotomodon alstoni mice (n = 20 and n = 20, respectively), and E, control and streptozotocin-treated rats (n = 40 and n = 40, respectively) in the 0.1–10 Hz range, under photopic conditions. ERG signals were normalized. Graphs show the average scalogram power ± s.e.m. throughout 1-minute recordings. The square brackets with the Roman numerals indicate the consistent peaks observed in each control condition. In the high-fat diet-induced obesity model, 0.1–0.8 (Low, L), 1–1.8 (M1, mid-low), and 2–4 Hz (M2, mid) bands were considered; 0.1–0.6 (I), 0.6–1 (II), and 1–1.7 (III) Hz bands in the spontaneous model of obesity, and 0.2–0.6 (I) and 0.6–2.5 (II) Hz bands in the type 1 diabetes model. AUC (0.1–10 Hz) and peak frequency analysis (P values were determined by unpaired Student’s t-test) of wavelet graphs in B, the high-fat diet-induced obesity D, the spontaneous model of obesity, and F, the type 1 diabetes models. Graphs show mean ± confidence interval. ROC curves and confusion matrix with performance measures for binary predictions (control vs. experimental cases) using the 0.1–10 Hz power spectra of G, control and high-fat diet-fed mice (n = 15 and n = 15, respectively), H, lean and spontaneously obese Neotomodon alstoni mice (n = 4 and n = 4, respectively), and I, control and streptozotocin-treated rats (n = 8 and n = 8, respectively). S, sensitivity; Sp., Specificity; NPV, negative predictive value; Ac., accuracy. J, Confusion matrix of multiclass prediction for the machine learning algorithm that discriminates between control and diseased rats at week 4, 6, 8, or 12 post-streptozotocin injection. Each column represents the instances in a predicted class, and rows represent the instances in an actual class. We used controls at week 4, 6, 8, and 12 (n = 8 at each time point).