Machine Learning Feasibility in Cochlear Implant Speech Perception Outcomes – Moving beyond Single Biomarkers for Cochlear Implant Performance Prediction
This study evaluates whether machine learning models can predict 6‑month post‑implantation speech perception in cochlear implant recipients using only preoperative clinical data. In a large cohort (n=1,877) drawn from the HERMES national CI database and a single institutional registry, XGBoost models outperformed traditional linear/logistic regression using Lazard features, reducing mean absolute error for CNC word scores (17.4% vs. 18.4%) and AzBio sentence scores (20.4% vs. 21.6%). More importantly, XGBoost demonstrated better calibration and discrimination at the lower performance quintiles (AUROC ~0.71 vs. ~0.59) and excelled at identifying individuals at risk for poor outcomes.
This represents the first preoperative stratification model for CI speech outcomes; while performance gains are modest and prospective validation is needed, ML shows promise in guiding counseling, rehabilitation planning, and personalized clinical trial design.