AI-Based Prediction of Bone Conduction Thresholds Using Air Conduction Audiometry Data.

Area of Science:

• Audiology
• Artificial Intelligence
• Machine Learning

Background:

• Pure tone audiometry (PTA) traditionally requires both air conduction (AC) and bone conduction (BC) measurements.
• Estimating BC thresholds and classifying air-bone gaps (ABGs) solely from AC data could streamline audiological assessments.

Purpose of the Study:

• To assess the feasibility of predicting BC thresholds and classifying ABG status using only AC data from PTA.
• To compare the performance of various machine learning models for these predictive tasks.

Main Methods:

• Utilized a large dataset of 60,718 PTA records from South Korea.
• Trained five machine learning models (DNN, LSTM, BiLSTM, RF, XGB) using AC thresholds, age, and sex as input features.
• Evaluated model performance using accuracy, sensitivity, precision, and F1 score with 5-fold cross-validation and SMOTE.

Main Results:

• LSTM and BiLSTM models showed superior performance in predicting BC thresholds, achieving approximately 60% accuracy within ±5 dB and 80% within ±10 dB.
• All models demonstrated better performance for ABG classification using a 10 dB criterion compared to a 15 dB criterion.
• Tree-based models (RF, XGB) yielded the highest classification accuracy (up to 0.512) and precision (up to 0.827).

Conclusions:

• Machine learning models can accurately predict BC thresholds and ABG status using only AC audiometry data.
• These findings advocate for the integration of AI tools in clinical audiology and telemedicine for remote hearing screening and diagnosis.
• Further clinical validation and implementation are recommended to enhance accessibility to hearing care services.