# Large language model–based prediction of speech intelligibility after Vibrant Soundbridge implantation using multidimensional outcome data: Part 2 of a prospective study

## TL;DR

This study uses a large language model to predict speech intelligibility outcomes after a specific hearing implant surgery, based on clinical data from 20 patients.

## Contribution

A novel predictive model using GPT-4o for estimating postoperative speech intelligibility in Vibrant Soundbridge implantation.

## Key findings

- A predictive model achieved an R² of 0.51 and r = 0.71 using BCPTA4, WRSmax, and age.
- Significant correlations were found between WRS65dB and BCPTA4, WRSmax, and age.
- The model provides a transparent tool for preoperative outcome estimation in VSB candidates.

## Abstract

Active middle ear implants (AMEIs) such as the Vibrant Soundbridge (VSB) offer an effective treatment option for patients with mixed or conductive hearing loss and large air–bone gaps, where conventional hearing rehabilitation often fails. However, postoperative outcomes—particularly speech intelligibility at 65 dB SPL in free-field (WRS65dB)—show high interindividual variability. This study aimed to develop a predictive model for WRS65dB based on four clinically relevant parameters: postoperative bone conduction thresholds (BCPTA4), unaided preoperative maximum speech intelligibility (WRSmax), Vibrogram threshold (VIBPTA4), and age. Data from 20 patients were analyzed. Spearman’s correlation revealed significant associations between WRS65dB and postoperative BCPTA4, preoperative WRSmax, and age. Using a seven-step approach supported by GPT-4o, we developed a sigmoid-transformed linear regression model. The final model included BCPTA4, WRSmax, and age and achieved an R² of 0.51, r = 0.71, RMSE = 6.18, and MAE = 4.67. Model performance was assessed by means of residual and outlier analysis. This model provides a transparent and clinically applicable tool for preoperative outcome estimation in VSB candidates. Further validation in larger, multicenter cohorts is needed to confirm its generalizability.

The online version contains supplementary material available at 10.1038/s41598-025-20919-5.