Interpretable machine learning with multimodal hearing data for diagnosing endolymphatic hydrops.

Area of Science:

• Audiology
• Medical Machine Learning
• Otolaryngology

Background:

• Endolymphatic hydrops (EH) diagnosis relies on clinical assessment, often lacking objective measures.
• Electrocochleography (ECochG) and pure-tone audiometry (PTA) provide auditory data but have limitations in diagnosing EH alone.

Purpose of the Study:

• To evaluate the diagnostic utility of ECochG for identifying EH.
• To enhance diagnostic accuracy by integrating ECochG and PTA data into machine learning (ML) models.
• To improve the clinical interpretability of ML models for EH diagnosis.

Main Methods:

• Prospective cohort study of 78 patients (156 ears) with suspected EH.
• ECochG and PTA examinations were conducted.
• Multimodal auditory data were used to develop and validate ML models, including Light Gradient Boosting Machine (LightGBM).

Main Results:

• ECochG-alone models showed limited diagnostic performance (AUC max 0.61).
• Integrating multimodal data into ML models substantially improved performance.
• The LightGBM model achieved an AUC of 0.92 and accuracy of 0.73 for predicting EH distribution.

Conclusions:

• Multimodal auditory data integrated into interpretable ML models significantly improve EH diagnostic accuracy.
• This approach offers an objective framework for early EH diagnosis and management.
• Further validation with larger, multi-center cohorts is warranted due to the study's limitations.