Multi-input Deep Learning Model for RP Diagnosis Using FVEP and Prior Knowledge

Retinitis Pigmentosa (RP) is a hereditary disease characterized by progressive damage to the visual pathway, ultimately leading to vision loss. Flash Visual Evoked Potential (FVEP) serves as an effective tool for diagnosing RP, and automatic classification of FVEP using deep learning can alleviate the workload of doctors and improve work efficiency. This study proposed a multi input neural network for RP and other anomaly recognition: MGPResNet. One branch of the model conducts full connection on manually crafted features to integrate them, while the other branch adopts a 1D ResNet as its basic architecture, it incorporates global convolutional blocks and pyramid pooling blocks to extract features from FVEP waveforms at deeper levels and different scales. Subsequently, the features extracted by the two branches are concatenated, followed by full connection and activation layers to output the classification probabilities. The model was validated on the FVEP datasets of two hospitals. The proposed method demonstrated excellent accuracy on clinical datasets, with an accuracy of 96.80%, average precision of 96.52%, average recall of 96.47%, and average F1_score of 96.49%. It validated the significant potential of deep learning in the analysis of visual electrophysiological signals, provided an important foundation and new insights for the future use of deep learning techniques in clinical diagnosis and treatment.