Empirical analysis of machine learning classifiers for retinal and optic nerve disorder detection using band-limited PERG features

Pattern Electroretinogram (PERG) signals offer critical insights into the functional status of retinal ganglion cells and are widely used for diagnosing optic nerve and retinal pathologies. While existing research has largely focused on binary classification or full-spectrum signal analysis, limited studies have addressed the use of frequency band-based feature extraction in multiclass classification. This study presents a novel methodology that segments PERG signals into defined frequency bands and extracts a robust set of 33 time-domain features per eye, encompassing statistical, morphological, and temporal characteristics. The dataset, originally comprising 15 diagnostic classes were consolidated into five clinically relevant pathology categories to enhance class balance and interpretability. Feature relevance was validated through one-way ANOVA, identifying key discriminative features (e.g., slope, std) with high F-values and p< 0.001. Several machine learning models were developed and evaluated, including Multilayer Neural Networks using the Levenberg–Marquardt algorithm (MLNN-LM), Random Forest, Bagged Trees, RUSBoost, and AdaBoost implemented via an Error-Correcting Output Codes (ECOC) framework. Among these, the AdaBoost-ECOC model demonstrated superior performance with an average classification accuracy of 99.11% ± 1.26, precision of 82.92% ± 0.59, and F1-score of 83.03% ± 0.55. These findings highlight the clinical applicability of band-specific time-domain features, supported by statistical validation, in enabling accurate and efficient multiclass classification of PERG signals.