# Development of a machine learning model for prognostic prediction of severe sudden sensorineural hearing loss with hyperbaric oxygen therapy

## TL;DR

A machine learning model was developed to predict recovery from severe sudden hearing loss, using factors like age, diabetes, and hyperbaric oxygen therapy sessions.

## Contribution

A novel machine learning model for predicting recovery in severe SSNHL, incorporating HBOT as a feature rather than assuming causality.

## Key findings

- The custom ML model achieved 89.36% test accuracy and an AUC of 0.8716.
- HBOT session count (≥10) was a key predictor in multiple models.
- Including HBOT data improved recovery prediction without asserting treatment efficacy.

## Abstract

Severe sudden sensorineural hearing loss (SSNHL) has heterogeneous causes and variable outcomes, making individualized prognosis difficult. We aimed to develop and evaluate a machine-learning (ML) model to predict recovery in severe SSNHL while treating hyperbaric oxygen therapy (HBOT) as an exposure feature rather than inferring causal treatment effects.

In a single-center retrospective cohort, we analyzed clinical and audiometric data from 231 in patients with severe SSNHL treated between January 2015 and January 2024. Recovery was defined by Siegel’s criteria; eligibility required ≥70 dB loss and treatment initiation within 1 month of onset. Candidate predictors included demographics, comorbidities, baseline thresholds, time to treatment, and HBOT variables (e.g., session count). We trained a custom multilayer perceptron with 12 input features and compared it with conventional algorithms. Performance was assessed using accuracy, F1 score, precision, recall, and area under the ROC curve (AUC).

Among 231 patients, the custom model achieved 89.36% test accuracy and an AUC of 0.8716, outperforming several conventional methods. Key predictors included age, diabetes, dizziness, and HBOT exposure. Notably, “≥10 HBOT sessions” showed high importance in logistic regression and SVM models, suggesting prognostic relevance of sufficient cumulative HBOT exposure.

Including HBOT information as a feature improved prediction of recovery in severe SSNHL; however, these findings do not establish the therapeutic efficacy of HBOT. The model may support clinician–patient decision-making by providing individualized recovery probabilities. Limitations include the retrospective single-center design, modest sample size, and class imbalance, underscoring the need for external validation and better adjustment for confounding.