Authors: Benjamin Burns¹; Michael Papazian²; Terrin Tamati¹; Xia Ning¹; Aaron Moberly²
¹The Ohio State University; ²Vanderbilt University Medical Center
Background: Speech recognition (SR) outcomes after cochlear implantation vary widely, motivating efforts to identify pre-operative (pre-op) variables predicting post-cochlear implant (CI) performance. Cognitive and linguistic (CL) abilities show promise as such predictors but remain largely unexplored. To determine if CL abilities improve post-CI SR prediction over conventional pre-op variables alone, we compared machine learning models with and without CL variables.
Method: Pre-op and 1-month post-CI SR measures (CNC words, AzBio sentences in quiet and noise) were collected from 42 adult postlingually deaf CI candidates with bilateral moderate-to-profound sensorineural hearing loss, along with demographic and pre-op CL variables. CL variables included Mini Mental State Examination (MMSE), Wide Range Achievement Test (WRAT) word reading, lexical and phonological access speed on Test of Word Reading Efficiency (TOWRE), nonverbal reasoning in Raven’s progressive matrices, and inhibition-concentration via visual Stroop color-word test. LASSO-regularized logistic regression models were fit with and without CL variables to classify post-CI SR into high, medium, and low performance groups.
Results: Most models incorporating CL variables demonstrated significantly improved predictions, with 21%, 72%, and 4% increases in macro-averaged AUROC for CNC bilateral best-aided, CNC CI-only, and AzBio in noise, respectively. The AzBio in quiet model did not benefit from including CL variables. CL feature selection frequency, as determined by LASSO, varied by outcome measure, with CNC models selecting CL variables 88% more frequently than AzBio models.
Conclusion: These findings suggest pre-op CL abilities contribute meaningful predictive value for early post-CI SR beyond conventional variables, particularly for word recognition. Considering CL abilities in pre-op CI candidacy evaluations may improve patient stratification and help set individualized expectations for post-CI SR.
