Introduction:

Accurate prediction of cochlear implant (CI) outcomes would improve patient counseling and clinical decision-making. We evaluated whether machine learning approaches could outperform a validated generalized linear regression model (GLM) that uses three key predictors (Hoppe et al., 2021 and 2023): preoperative maximum word recognition score (WRSMax), aided word recognition score at 65 dB SPL (WRS65(HA)), and age at implantation.

Methods:

We retrospectively analyzed data from 229 adult CI recipients (191 unilateral, 38 bilateral) with post-lingual deafness and WRSMax >0. Using WRSMax, WRS65(HA), and age at implantation as predictors, and WRS65 with CI as the outcome variable, we compared machine learning approaches (regularized Elastic Net regression, Generalized Additive Model (GAM), eXtreme Gradient Boosting (XGBoost), Random Forest decision tree ensemble, and their mean ensemble) with the validated GLM through 100 bootstrap iterations.

Results:

Results showed limited predictive performance across all models. GLM achieved Root Mean Square Error (RMSE) of 23.5 percentage points (95%-CI: 20.6-26.3) and R² of 3.1% (CI: 0.8-7.7%). Machine learning models performed similarly: Elastic Net (RMSE: 23.4, CI: 20.8-25.8; R²: 0.8%, CI: 0-1.2%), GAM (RMSE: 23.9, CI: 20.8-26.5; R²: 1.0%, CI: 0-1.2%), XGBoost (RMSE: 25.1, CI: 22.6-27.4; R²: 0.7%, CI: 0-0.8%), Random Forest (RMSE: 25.3, CI: 22.1-28.0; R²: 0.5%, CI: 0-0.7%), and mean ensemble (RMSE: 23.6, CI: 20.7-26.0; R²: 1.1%, CI: 0-1.1%). Feature engineering attempts yielded no improvements.

Discussion:

Advanced modeling approaches did not outperform the simpler GLM. The consistently low R² values suggest limited predictive power of the three previously suggested key variables. Future research should investigate additional predictors such as cognitive measures, auditory processing, and electrode position characteristics.

R.B. received travel cost compensation and course fees for surgical training from Cochlear and MedEL

R.S. received travel cost compensation and course fees for surgical training from MedEL

A.R. received travel cost compensation from MedEL