Towards AI-Based Strep Throat Detection and Interpretation for Remote Australian Indigenous Communities

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Public Health

Background:

• Streptococcus pharyngitis (strep throat) is a significant health issue in remote Australian Indigenous communities with limited medical access.
• Delayed diagnosis and treatment increase risks of severe complications like acute rheumatic fever and rheumatic heart disease.

Purpose of the Study:

• To develop and validate a proof-of-concept AI diagnostic model for strep throat detection.
• To support clinicians in underserved areas with timely and accurate diagnoses.
• To enhance healthcare accessibility in resource-constrained settings.

Main Methods:

• A lightweight Swin Transformer-based image classifier was developed to predict strep throat from throat images.
• An explainable BLIP-2-based image annotation system was integrated to identify clinical features (tonsillar swelling, erythema, exudate).
• Synthetic labels for explainability were generated using GPT-4o-mini, and the model was optimized for low-resource environments.

Main Results:

• The AI classifier achieved 97.1% accuracy and an ROC-AUC of 0.993.
• The model demonstrated a fast inference time of 13.8 ms and a compact size of 28 million parameters.
• These metrics indicate suitability for deployment in resource-limited settings.

Conclusions:

• The AI-based diagnostic model shows significant potential for improving healthcare access and outcomes in remote and underserved regions.
• This proof-of-concept illustrates the value of AI-assisted diagnostics in clinical decision-making.
• Further research can build upon this model to address healthcare disparities globally.