Machine Learning Feasibility in Cochlear Implant Speech Perception Outcomes-Moving Beyond Single Biomarkers for Cochlear Implant Performance Prediction.

This was a retrospective registry analysis, not a prospective clinical validation. Over half of the initially compiled patients were removed because key outcome or preoperative speech data were missing, and missing data may not have been random. The features used were limited to those available before implantation in both datasets, so cognition, comorbidities, electrode placement, and other potential factors were not included. The performance categories were based on quintiles and author consensus, not published consensus guidelines for good or poor cochlear implant performance. The dataset was biased toward a smaller number of higher-volume cochlear implant centers, and the models were not externally validated. Probability filtering improved some metrics, but the model could be applied to only half of the evaluated patients.

Funding was reported from the National Institutes of Health and the Foundation for Barnes-Jewish Hospital.

Do not read this as a ready clinical tool. The numerical prediction advantage was less than 1%, and the lowest ΔCNC quintile was correctly named only 42% of the time. The authors state that external validation is needed before implementation and that the model should never be used to deny a cochlear implant.