ERG Classification by Using ML Methods Based on Short-Time Fourier Transform

Electroretinography (ERG) is a noninvasive method for evaluating the functional status of the retina by using light stimulation. This method has demonstrated its efficacy in generating valuable features for ERG classification, as evidenced by existing research findings. This study uses the Short-Time Fourier Transform, a time-frequency domain method, to extract features from ERG signals, which are then used to train Decision Tree models for the classification of these signals. The main objective of the study is to compare different window functions, sizes and overlaps in order to determine which combination yields the most optimal features for the classification of ERG signal. The dataset used for this investigation is from a large database of electroretinogram signals consisting of 5 types of signals including Maximum 2.0 ERG response, Photopic 2.0 ERG response, Scotopic 2.0 ERG response, oscillatory Potentials, and Photopic 2.0 EGR Flicker response. The results from this investigation have indicated that this methodology has considerable potential, but also presents several challenges, such as imbalanced datasets and the resolution trade-off. This work has the potential to contribute to the development of more effective tools to assess the functional health of the retina, enhancing the diagnosis and treatment of various retinal diseases.