Predicting the hearing outcome in sudden sensorineural hearing loss via machine learning models

Area of Science:

• Otolaryngology
• Medical Informatics
• Machine Learning

Background:

• Sudden sensorineural hearing loss (SSHL) presents high heterogeneity, leading to variable patient outcomes.
• Predictive models are needed to guide clinical management and prognostication for SSHL patients.

Purpose of the Study:

• To develop and compare machine learning models for predicting hearing outcomes in SSHL.
• To identify the most effective model for clinical application in SSHL prognosis.

Main Methods:

• A retrospective study included 1220 in-patient SSHL cases.
• Four machine learning methods were developed: deep belief network (DBN), logistic regression (LR), support vector machine (SVM), and multilayer perceptron (MLP).
• Models were trained using six feature collections derived from 149 potential predictors, evaluating accuracy, precision, recall, F-score, and ROC-AUC.

Main Results:

• The DBN model achieved the highest predictive accuracy (77.58%) and ROC-AUC (0.84) with 149 variables.
• DBN performance decreased after feature pruning.
• LR, SVM, and MLP models showed peak performance with only three variables (ROC-AUC 0.79-0.81), with performance declining as more features were added.

Conclusions:

• Deep belief networks (DBN) offer robust prediction for SSHL when provided with sufficient features.
• Logistic regression (LR) using three readily available variables (time to treatment, initial hearing level, audiogram) is more practical for early clinical prediction of SSHL outcomes.