AGCS – The AudioGene Confidence Score: an interpretable post-hoc confidence metric for machine learning-based gene prediction in autosomal dominant non-syndromic hearing loss

Hearing Loss affects millions of people worldwide, and for many patients the underlying cause is genetic. Identifying the specific gene responsible for hearing loss is critical for accurate diagnosis, family counseling, and guiding further genetic testing. However, dozens of different genes can cause hearing loss that looks nearly identical on a standard hearing test, making it hard to determine which gene is responsible from hearing test data alone. AudioGene is a computer program developed at the University of Iowa that uses artificial intelligence to analyze a patient’s hearing test results and predict the most likely causative gene. While AudioGene produces a ranked list of candidate genes, clinicians previously had no way of knowing whether a given prediction was reliable or not. Every prediction looked the same regardless of how confident the model was. This thesis introduces the AudioGene Confidence Score (AGCS), a tool that assigns a confidence level to each AudioGene prediction. Predictions are placed into one of five tiers, ranging from Extremely Low to Extremely High confidence, giving clinicians a clear signal for when to trust the model’s output and when to seek additional evidence before making a diagnostic decision. Predictions in the highest confidence tiers were correct more than 91% and 95% of the time. As a secondary contribution, this work rebalanced AudioGene’s training data to better reflect the true frequency of each gene in the real patient population. Real patient case studies demonstrate how the confidence score can guide clinical decision-making in practice.