AutoAudio: Deep Learning for Automatic Audiogram Interpretation

Matthew G. Crowson, Jong Wook Lee, Amr F. Hamour, Rafid Mahmood, Aaron Babier, Vincent Lin, Debara L. Tucci, Timothy C. Y. Chan

Abstract

Hearing loss is a leading cause of disability worldwide, yet access to diagnostic interpretation remains a significant bottleneck in global healthcare. This study presents AutoAudio, a deep learning proof-of-concept designed to automatically interpret diagnostic audiograms by treating them as visual data. Using a dataset of 1,007 adult audiogram reports, the researchers evaluated several Residual Network (ResNet) architectures. The ResNet-101 model achieved a peak classification accuracy of 97.5% across four primary categories: normal, conductive, sensorineural, and mixed hearing loss. These results suggest that computer vision can effectively automate the interpretation of legacy and modern audiometric plots, potentially expanding access to hearing care in underserved regions.

Introduction: The Global Audiology Gap

Hearing loss is the most prevalent sensory deficit in humans and a major contributor to years lived with disability. Beyond the immediate sensory impact, untreated hearing loss is linked to social isolation, cognitive decline, and reduced quality of life. As the global population ages, the burden of hearing loss is forecasted to rise sharply, yet the audiology workforce is struggling to keep pace.

In many developing nations, the ratio of audiologists to the population is as low as one per million. Even in developed nations like the United States, the demand for hearing evaluations is expected to outstrip supply by 50% in the coming years. While portable testing hardware has become more accessible, the interpretation of the resulting audiograms—the graphs mapping hearing thresholds across frequencies—still requires specialized expertise. AutoAudio was developed to bridge this gap by automating the diagnostic interpretation of these visual records.

Methodology: A Computer Vision Approach

Unlike previous automation attempts that rely on raw digital data points, AutoAudio utilizes a computer vision framework. This approach allows the model to interpret traditional "plot-based" audiograms, making it compatible with both modern digital outputs and scanned historical paper records.

Model Architecture and Training

The researchers employed Transfer Learning, utilizing four variants of the Residual Network (ResNet) architecture: ResNet-32, 50, 101, and 152. These models were originally pre-trained on the ImageNet database. However, because audiograms differ significantly from natural images, the team re-trained every layer of the networks to specialize in medical plot recognition.

To improve the model's robustness and simulate real-world conditions (such as slightly crooked scans or varying image quality), the team used image augmentation techniques, including rotation, warping, and contrast adjustments. This artificially expanded the training set and helped the model generalize better to unseen data.

Explainability with Grad-CAM

To ensure the model was making decisions based on clinically relevant features rather than artifacts, the researchers used Grad-CAM (Gradient-weighted Class Activation Mapping). This technique generates heat-maps over the audiogram images, highlighting the specific visual regions—such as the gap between air and bone conduction lines—that most influenced the model’s classification.

Dataset and Experimental Design

The study utilized a convenience sample of 1,007 audiogram reports from adult patients at the Sunnybrook Health Sciences Center (2017–2019). The dataset was categorized by expert audiologists into four morphologies:

  • Sensorineural loss (435 cases)
  • Mixed loss (214 cases)
  • Normal hearing (207 cases)
  • Conductive loss (151 cases)

The data was split into an 80% training set (806 images) and a 20% "hold-out" validation set (201 images) to evaluate performance on entirely new data.

Key Results and Discussion

High-Accuracy Classification

The ResNet-101 architecture emerged as the top performer, achieving a classification accuracy of 97.5%. Interestingly, the study found diminishing returns with increased network depth; the performance difference between the 32-layer and 152-layer versions was less than 2%. This suggests that "lighter" models may be sufficient for clinical deployment, offering faster training and inference times without significant loss in accuracy.

Clinical Logic through Pattern Recognition

The Grad-CAM analysis revealed that the model independently "learned" fundamental clinical principles. For example, it focused heavily on the "air-bone gap" to distinguish between conductive and sensorineural losses. This indicates that the deep learning model was not just memorizing data but was identifying the spatial relationships that human experts use for diagnosis.

The Challenge of Mixed Hearing Loss

The most common misclassifications occurred with mixed hearing loss. Clinically, mixed loss is the most complex morphology, as it requires the simultaneous identification of thresholds below the normal range and a significant gap between air and bone conduction. The model occasionally failed to recognize subtle bone conduction thresholds, leading it to misclassify mixed losses as purely conductive.

Limitations and Future Directions

While AutoAudio represents a significant step forward, several hurdles remain before clinical implementation:

  1. Generalizability: The model was trained on standardized plots from a single institution. Future work must test its performance on varied formats, different scales, and combined-ear graphs.
  2. Clinical Context: Diagnosis involves more than just audiogram morphology. A full clinical picture requires patient history, tympanometry, and word recognition scores, which the current model does not incorporate.
  3. Human vs. Machine Comparison: The next phase of research will involve direct head-to-head testing between AutoAudio and practicing audiologists to establish inter-rater reliability.

Conclusion

AutoAudio demonstrates that deep learning can accurately interpret the visual language of hearing loss. By providing a "backwards compatible" tool that can digitize and triage legacy paper records as easily as modern digital plots, this technology could significantly enhance access to hearing care in resource-limited settings worldwide.

Summary updated

Abstract

Abstract Objectives Hearing loss is the leading human sensory system loss, and one of the leading causes for years lived with disability with significant effects on quality of life, social isolation, and overall health. Coupled with a forecast of increased hearing loss burden worldwide, national and international health organizations have urgently recommended that access to hearing evaluation be expanded to meet demand. Methods The objective of this study was to develop ‘AutoAudio’ – a novel deep learning proof-of-concept model that accurately and quickly interprets diagnostic audiograms. Adult audiogram reports representing normal, conductive, mixed and sensorineural morphologies were used to train different neural network architectures. Image augmentation techniques were used to increase the training image set size. Classification accuracy on a separate test set was used to assess model performance. Results The architecture with the highest out-of-training set accuracy was ResNet-101 at 97.5%. Neural network training time varied between 2 to 7 hours depending on the depth of the neural network architecture. Each neural network architecture produced misclassifications that arose from failures of the model to correctly label the audiogram with the appropriate hearing loss type. The most commonly misclassified hearing loss type were mixed losses. Conclusion Re-engineering the process of hearing testing with a machine learning innovation may help enhance access to the growing worldwide population that is expected to require audiologist services. Our results suggest that deep learning may be a transformative technology that enables automatic and accurate audiogram interpretation.
BibTeX
@article{crowson2020autoaudio,
  author = {Matthew G. Crowson and Jong Wook Lee and Amr F. Hamour and Rafid Mahmood and Aaron Babier and Vincent Lin and Debara L. Tucci and Timothy C. Y. Chan},
  title = {{AutoAudio: Deep Learning for Automatic Audiogram Interpretation}},
  journal = {J. Medical Syst.},
  year = {2020},
  doi = {10.1007/s10916-020-01627-1},
}
Figure