Candidacy for Cochlear implantation: Validating a novel Cochlear implant candidacy calculator against gold‐standard, in‐clinic audiometric assessments

Abstract

Objectives: Cochlear implants (CI) are reliable implantable devices that are highly cost-effective in reducing the burden of hearing loss at an individual and societal scale. However, only 10% of CI candidates are aware of their candidacy and receive a CI. A web-based screening tool to assess CI candidacy may make many more individuals aware of their candidacy for cochlear implantation. The objective of this study was to validate and optimize the online Cochlear Implant Candidacy Calculator against in-clinic audiometric testing.Methods: Audiogram data and word discrimination scores for 132 patients who underwent initial CI consultation at the Johns Hopkins Cochlear Implant Center in 2020 were inputted into the calculator. Candidacy results from the calculator were compared against formal clinical diagnoses provided by the audiologist at the time of visit. Receiver Operating Characteristic (ROC) and Area Under the Curve (AUC) analyses were performed to identify optimal diagnostic thresholds.Results: Of the resulting 132 patients, 54 presented with single-sided deafness (SSD), and 114 were clinically determined to be CI candidates. ROC AUC analyses identified optimal thresholds of high-frequency PTA ≥65 dB and word discrimination score ≤ 50%.To maximize sensitivity at the expense of specificity, diagnostic thresholds of high-frequency PTA ≥ 65 dB and word discrimination score ≤ 70% were chosen, which yielded accuracy, sensitivity, specificity, and ROC AUC of 0.90, 0.94, 0.82, and 0.88, respectively. Conclusion:The novel online CI Candidacy Calculator exhibits high sensitivity and accuracy, and moderate specificity. The calculator may thereby be useful in increasing awareness of potential CI candidacy, increasing prevalence of CIs, and decreasing the burden of hearing loss.

Machine Learning Approach for Screening Cochlear Implant Candidates: Comparing With the 60/60 Guideline

“…Compared with the 60/ 60 guideline, the RF model proposed in this study attained higher sensitivity (0.91 versus 0.96), positive predictive value (0.97 versus 1.00), and accuracy (0.89 versus 0.96) while considerably improving specificity (0.42 versus 1.00) and negative predictive value (0.16 versus 0.53), with similar results in aggregate across the bootstrapped model fits. Furthermore, the RF model yielded an area under the receiver operating characteristic curve of 0.96, which was higher than recent referral tools that ranged from 0.76 to 0.88 for AUC (24,25). Our proposed machine learningbased model for CICE referrals thus minimizes false positives and false negatives and demonstrates generalizability after cross-validation.…”

“…Using the largest cohort to date, this study is the first to create a novel machine learning-based referral algorithm, incorporating both demographic and audiometric factors, that can predict CI candidacy under Medicare-eligible, traditional, and expanded criteria. Other referral guidelines and candidacy calculators are solely based on audiometry and for patients who would primarily be eligible under Medicare and traditional criteria (23)(24)(25)(26)(27)(28)(29). Assessment of five CICE referral tools demonstrated best performance in balancing sensitivity and specificity by the 60/60 guideline, although all screening tools performed worse on a validation cohort (30).…”

“…Assessment of five CICE referral tools demonstrated best performance in balancing sensitivity and specificity by the 60/60 guideline, although all screening tools performed worse on a validation cohort (30). More recent screening tools, which also only incorporate audiometry, used smaller cohorts of 132 (25) and 252 (24) patients with no cross- or external-validation of the proposed guidelines.…”

“…Previous referral guidelines, such as 60/60 presented by Zwolan et al (23) and 65/70 presented by So et al (25), used cutoffs for PTA and WRS, suggesting that patients with a PTA of at least 60 or 65 dB HL and WRS of 60% or 70% and less should undergo CICE. To improve upon these guidelines and allow for incorporation of expanding candidacy criteria, a random forest (RF) classification model was trained on 772 adult patients undergoing CICE at our institution between 2015 and 2020.…”

“…Recently, multiple screening tools have been developed using data from standard audiograms to assist hearing professionals with their clinical decision making (23)(24)(25)(26)(27)(28)(29). These tools have inconsistent sensitivity and specificity, continue to propose complex guidelines, and are only modestly successful at identifying potential CI candidates with an external data set (30).…”