Feasibility of Random Forest Classifier Machine Learning Algorithm as Initial Screening for Cochlear Implantation Candidacy

Introduction: Candidacy for cochlear implantation (CI) is primarily based on aided speech audiometry scores; however, the standard audiogram includes only pure tone thresholds and unaided word recognition scores. Selecting which patients are referred for a evaluation is often decided by the clinician’s gestalt or by clinical rules such as the 60/60 rule (word recognition score of <60% and pure tone average in the better ear of ≥60dB). The objective of this work is to evaluate the accuracy of a machine learning algorithm in predicting aided speech scores from standard audiometry.

Methods: A retrospective review of an institutional database identified patients who underwent both a screening audiogram and CI evaluation with AzBio speech testing bilaterally. Each ear was considered separately, yielding a total of 417 observations from 229 patients. A random forest classifier (RFC) was trained to predict an aided AzBio score <50% using 5-fold cross validation on 90% of the dataset with 10% reserved for evaluation. The aided AzBio threshold was chosen as the highest score that would generally be considered for CI. Linear regression and the 60/60 rule were also evaluated.

Results: The RFC achieved an accuracy of 92.9%, specificity of 66.7%, and a sensitivity of 100%, with an area under the curve (AUC) of 83.3%. Linear regression achieved an accuracy of 66.7%, specificity of 22.2%, sensitivity of 78.8%, and AUC of 50.5%. The 60/60 rule achieved an accuracy of 76.2%, specificity of 100% and a sensitivity of 69.7%.

Conclusions: The RFC outperformed the linear regression model and the 60/60 rule based approach in predicting an aided AzBio <50% from a standard audiogram. The high sensitivity achieved by the RFC suggests that this method was able to capture all patients that meet the aided threshold for CI candidacy, while the 60/60 rule had a false negative rate of 30.3% on the same cohort. This suggests that the RFC may be a better screening test for CI candidacy. Further study on a larger cohort is required to confirm and integrate these findings into clinical practice.