AutoAudio: Deep Learning for Automatic Audiogram Interpretation.

Area of Science:

• Audiology and Artificial Intelligence
• Medical Diagnostics

Background:

• Hearing loss is a major global health issue, impacting quality of life and increasing disability.
• There is an urgent need to expand access to hearing evaluations worldwide due to rising hearing loss prevalence.

Purpose of the Study:

• To develop and evaluate AutoAudio, a novel deep learning model for rapid and accurate interpretation of diagnostic audiograms.
• To assess the performance of different neural network architectures in classifying hearing loss types.

Main Methods:

• Trained various neural network architectures using adult audiogram reports (normal, conductive, mixed, sensorineural).
• Employed image augmentation techniques to expand the training dataset.
• Evaluated model performance using classification accuracy on a separate test set.

Main Results:

• The ResNet-101 architecture achieved the highest accuracy at 97.5% on the test set.
• Neural network training times ranged from 2 to 7 hours.
• Mixed hearing loss types were the most commonly misclassified.

Conclusions:

• Deep learning models like AutoAudio show potential for automatic and accurate audiogram interpretation.
• Machine learning innovations can help re-engineer hearing testing processes to meet increasing global demand for audiology services.