Predicting speech discrimination scores from pure-tone thresholds-A machine learning-based approach using data from...

Area of Science:

• Audiology
• Machine Learning
• Medical Diagnostics

Background:

• Accurate hearing loss assessment is crucial for patient treatment and rehabilitation.
• Current metrics like pure-tone audiometry (PTA) and speech discrimination scores (SDS) can be influenced by human factors, necessitating cross-validation.
• Identifying potential malingering in hearing tests is important for accurate diagnosis.

Purpose of the Study:

• To develop a machine learning model for predicting SDS based on PTA thresholds.
• To enhance the accuracy and reliability of hearing loss diagnostics.
• To create a tool for identifying potential malingering in patients undergoing hearing tests.

Main Methods:

• A Random Forest-based machine learning approach was utilized.
• A large-scale dataset of 12,697 subjects was collected for model training and evaluation.
• The model's performance was compared against other machine learning algorithms like Support Vector Machine and Multi-layer Perceptron.

Main Results:

• The Random Forest model achieved high prediction accuracy for SDS: 95.05% for the left ear and 96.64% for the right ear.
• The proposed model demonstrated superior performance compared to other tested machine learning algorithms.
• The study validates the effectiveness of using PTA thresholds to predict SDS.

Conclusions:

• The developed Random Forest model offers a reliable and accurate method for estimating SDS from PTA thresholds.
• This approach can serve as a practical screening tool to identify potential malingering in hearing loss evaluations.
• The findings support the integration of machine learning in audiological diagnostics for improved patient care.