Paper Title
A IMAGE DEEP LEARNING MODEL PREDICTS THE RECOVERY RATE OF SUDDEN HEARING LOSS

Abstract
Background: The prognostic factors for sudden sensorineural hearing loss are related to the patient's comorbidities, the time from disease onset to treatment, the degree of hearing loss, the patient's age, and the presence of dizziness. The shape of the patient's initial audiogram plays an important role in the patient's prognosis. The purpose of this study is to use deep learning to build a model to analyze the initial audiograms of patients with sudden hearing loss on the prognosis of our hospital. Objective: Develop a deep learning model to analyze the initial audiograms of patients with sudden hearing loss on the prognosis. Methods: From January 2021 to August 2023, a total of 135 patients diagnosed with sudden sensorineural hearing loss and hospitalized in our hospital were collected their pure tone audiograms. Using Python programming language and Pytorch deep learning database, a model was trained to predict the patient's prognosis by using the ResNet50 neural convolution network module with the patient's initial pure tone audiogram as input data. Results: Pure tone audiometry data of 135 patients was pooled into ResNet50 neural convolution network to build the model. On the fifth day of hospitalization, if the patients were divided into two groups: no recovery and recovery, the accuracy of the validation group of the trained model could reach 88.37%; if the patients were divided into four groups: no recovery, mild recovery, partial recovery, and healed, the accuracy of the training group could reach 100%, and the accuracy of the validation group was 40%. Conclusion: The number of patients collected in this study was small in terms of the training of the deep learning module, but the prediction rate of whether there was a response to treatment on the fifth day of hospitalization was as high as 88% for patients who started hospitalization, but if the patients were further divided into four groups, the amount of data in each group would be even lower, and the prediction rate would be lower. It is expected that if the amount of training data can be further increased, a high accuracy prediction module can be established.