# Machine Learning Versus Simple Clinical Models for Cochlear Implant Outcome Prediction

## TL;DR

This study compares machine learning and simple models to predict cochlear implant outcomes, finding that both perform similarly and only slightly better than a basic model.

## Contribution

The study evaluates the predictive performance of various statistical and machine learning models against a Null model for cochlear implant outcomes.

## Key findings

- All models showed similar predictive accuracy with root mean squared errors between 26.28 and 30.74 percentage points.
- Simpler models performed nearly as well as complex machine learning models in predicting cochlear implant outcomes.
- Pre-implantation clinical variables had limited predictive validity for cochlear implant outcomes.

## Abstract

Background/Objectives: Cochlear implantation is the most widely used treatment option for patients with severe to profound hearing loss. Despite being a relatively standardized surgical procedure, cochlear implant (CI) outcomes vary considerably among patients. Several studies have attempted to develop predictive models for CI outcomes but achieving accurate and generalizable predictions remains challenging. The present study aimed to evaluate whether simple and complex statistical and machine learning models could outperform the Null model based on various pre-CI implantation variables. Methods: We conducted a retrospective analysis of 236 ears with postlingual profound sensorineural hearing loss (SNHL) and measurable residual hearing (WRSmax > 0%) at the time of implantation. The median postoperative word recognition score with CI (WRS65(CI)) was 75% [Q1: 55%, Q3: 80%]. The dataset was divided using a 70:15:15 split into training (n = 165), validation (n = 35) and test (n = 36) cohorts. We evaluated multiple modeling approaches: different Generalized Linear Model (GLM) approaches, Elastic Net, XGBoost, Random Forest, ensemble methods, and a Null model baseline. Results: All models demonstrated similar predictive performance, with root mean squared errors ranging from 26.28 percentage points (pp) to 30.74 and mean absolute errors ranging from 20.62 pp to 23.75 pp. Coefficients of determination (R2) ranged from −0.468 to −0.073. Bland–Altman analyses revealed wide limits of agreement and consistent negative bias, while Passing–Bablok regression indicated calibration errors. Nonetheless, all models incorporating predictors significantly outperformed the Null model. Conclusions: Increasing model complexity yielded only marginal improvements in predictive accuracy compared with simpler statistical models. Pre-implantation clinical variables showed limited evidence of predictive validity for CI outcomes, although further research is needed.