Spontaneous ERG oscillations help predict the disease in patients with early risk factors of DR.

A, Illustrative signals from spontaneous ERGs of control subjects and patients with overweight (OW), obesity, metabolic syndrome (MetS), and diabetes but no DR (DM no DR) under photopic conditions. Signals are normalized. Wavelet analysis for B, each group separately and C, combining all unhealthy conditions into one disease group. Control group is defined in Methods. Graphs show the average scalogram power ± s.e.m. throughout 20-second recordings between 0.3–40 Hz (host graph) and 0.3–2 (I), 10–20 (II), and 20–40 (III) Hz (inset graphs), where consistent peaks were observed in the control group. D, AUC and peak frequency analysis in the I, II, and III bands of wavelet power spectra from control and disease groups (P values were determined by unpaired Student’s t-test). Graphs show mean ± confidence interval. E, ROC curve and F, confusion matrix with performance measures for the Random Forest model discriminating control from disease cases in control and disease groups, using the 0.3–40 Hz power spectra. G, ROC curve and H, confusion matrix with performance measures corresponding to the validation of our predictive model thanks to an external and independent validation dataset. I, Classification performance of our model versus experts, when experts have access to full eye examination alone or combined with reference tests to diagnose diabetes (fasting glycemia, creatinine, triglycerides, and cholesterol) or full laboratory (glycemia, HbA1c levels, HOMA-I, and lipid profile) and non-laboratory (blood pressure, weight and waist circumference, and BMI) tests. Precision, recall, and F1-score are reported. I, ROC curve and J, confusion matrix with performance measures corresponding to the performance of our predictive model using ERG-derived power spectra recorded by only one ERG device. In A-F and I, control (metabolically healthy, n = 80) and disease (OW, n = 40; obese, n = 14; MetS, n = 66; and diabetes with no DR, n = 68) groups. In G, H, control (metabolically healthy, n = 11) and disease (OW, n = 5; obese, n = 3; MetS, n = 3; and diabetes with no DR, n = 3) groups. In J, K, control (metabolically healthy, n = 37) and disease (OW, n = 18; obese, n = 8; MetS, n = 42; and diabetes with no DR, n = 33) groups. S, sensitivity; Sp., Specificity; NPV, negative predictive value; Ac., accuracy.