A Deep Learning Model to Predict Conductive Hearing Loss in Aboriginal and Torres Strait Islander Children Using Otoscopy
Dr Rahim Habib, University of Sydney, Australia
Authors List
Introduction
Long-term conductive hearing loss (CHL) can significantly impact speech and language development, academic performance, future employment opportunities and quality of life in Aboriginal and Torres Strait Islander children. In rural and remote areas, access to otolaryngologists and audiologists is limited. Artificial intelligence (AI) algorithms may be developed to predict CHL using otoscopic images.
Aims
To determine if an AI, deep learning model can predict the presence and degree of CHL in Aboriginal and Torres Strait Islander children using otoscopic images alone.
Methods
Clinical data were collected from the Deadly Ears Program in Queensland, Australia from 2017 to 2020. Otoscopic images and audiometric clinical data were ascertained from face-to-face examinations conducted by nurses and audiologists. The degree of hearing loss was classified by audiologists as the ground-truth (normal, mild, moderate, and moderately severe to profound hearing loss). Deep learning and transfer learning methods were used to develop a model to predict the degree of hearing loss by reviewing otoscopic images. Performance was evaluated by accuracy, precision, and recall.
Results
4436 otoscopic images with corresponding audiometry results were used for training and achieved 87.0% accuracy, 87.4% precision and 86.4% recall. In a test cohort of 69 otoscopic images not used for training, the model achieved 88.4% accuracy to predict the presence or absence of hearing loss and 78.2% accuracy to correctly predict the degree of hearing loss. In test cases with middle ear effusions, the model correctly identified the likelihood of hearing loss in 17 of 21 (80.9%) presentations.
Conclusion
Deep learning models can predict the presence and severity of hearing loss in Aboriginal and Torres Strait Islander children using otoscopic images alone. AI has the potential to improve triage processes by accelerating the recognition of children at-risk for hearing loss in rural and remote areas.
- Habib, Al-Rahim, University of Sydney, Sydney, New South Wales, Australia
- Perry, Chris, University of Queensland, Brisbane, Queensland, Australia
- Crossland, Graeme, Royal Darwin Hospital, Darwin, Northern Territory, Australia
- Patel, Hemi., Royal Darwin Hospital, Darwin, Northern Territory, Australia
- Kong, Kelvin, University of Newcastle, Newcastle, New South Wales, Australia
- Gunasekera, Hasantha, University of Sydney, Sydney, New South Wales, Australia
- Sacks, Raymond, University of Sydney, Sydney, New South Wales, Australia
- Kumar, Ashnil, University of Sydney, Sydney, New South Wales, Australia
- Singh, Narinder, University of Sydney, Sydney, New South Wales, Australia
Introduction
Long-term conductive hearing loss (CHL) can significantly impact speech and language development, academic performance, future employment opportunities and quality of life in Aboriginal and Torres Strait Islander children. In rural and remote areas, access to otolaryngologists and audiologists is limited. Artificial intelligence (AI) algorithms may be developed to predict CHL using otoscopic images.
Aims
To determine if an AI, deep learning model can predict the presence and degree of CHL in Aboriginal and Torres Strait Islander children using otoscopic images alone.
Methods
Clinical data were collected from the Deadly Ears Program in Queensland, Australia from 2017 to 2020. Otoscopic images and audiometric clinical data were ascertained from face-to-face examinations conducted by nurses and audiologists. The degree of hearing loss was classified by audiologists as the ground-truth (normal, mild, moderate, and moderately severe to profound hearing loss). Deep learning and transfer learning methods were used to develop a model to predict the degree of hearing loss by reviewing otoscopic images. Performance was evaluated by accuracy, precision, and recall.
Results
4436 otoscopic images with corresponding audiometry results were used for training and achieved 87.0% accuracy, 87.4% precision and 86.4% recall. In a test cohort of 69 otoscopic images not used for training, the model achieved 88.4% accuracy to predict the presence or absence of hearing loss and 78.2% accuracy to correctly predict the degree of hearing loss. In test cases with middle ear effusions, the model correctly identified the likelihood of hearing loss in 17 of 21 (80.9%) presentations.
Conclusion
Deep learning models can predict the presence and severity of hearing loss in Aboriginal and Torres Strait Islander children using otoscopic images alone. AI has the potential to improve triage processes by accelerating the recognition of children at-risk for hearing loss in rural and remote areas.
