Authors: Marta Campi¹; Alexander Huber²; Tobias Goehring¹
¹Deep Hearing Lab, University of Zurich; ²University Hospital Zurich, University of Zurich
Background: The prediction of individual speech outcomes after cochlear implantation remains challenging, with reported explained variances of R2 = 0.07 to 0.27 and prediction errors too large for clinical counselling. Whether this reflects insufficient modelling capability or a fundamental limit of the available predictor space is unknown. Here we compare statistical models across prediction tasks and discuss their clinical relevance.
Method: We performed a retrospective analysis of a single-centre adult cohort (538 adults) from the Swiss National Cochlear Implant Database using linear models, ensemble methods, and a state-of-the-art foundation model for tabular data, with varying feature sets and imputation strategies under cross-validation. We compared continuous score prediction, binary classification (poor outcome, decline), and unsupervised profiling of post-operative speech patterns across five speech tests.
Results: No modelling approach achieved clinically useful prediction of individual post-operative scores (cross-validated R2 below 0.03). Patients across hearing-loss severity converge to overlapping post-operative ranges, limiting the variance available for regression. Binary classification – identifying patients performing below a clinical threshold or declining after implantation – proved more successful (AUC = 0.67 and 0.77 respectively), with onset type, articulation, contralateral hearing, and duration of hearing loss as consistent predictors. Unsupervised clustering identified distinct performance subgroups, including patients with atypical patterns across speech dimensions, but these could not be predicted from pre-operative data alone.
Conclusion: The limitation appears informational rather than algorithmic. We propose alternatives to score prediction: pre-operative risk screening, post-operative profiling to guide rehabilitation, and early post-operative monitoring to identify patients requiring intervention.
