Imputation of missing values for cochlear implant candidate audiometric data and potential applications.

Area of Science:

• Audiology
• Data Science
• Medical Informatics

Background:

• Cochlear implant (CI) candidacy relies on comprehensive audiometric data.
• Missing data in audiograms can limit patient selection and research.
• Developing robust imputation methods is crucial for maximizing clinical datasets.

Purpose of the Study:

• To evaluate the real-world performance of various imputation algorithms on CI candidate audiometric data.
• To determine the efficacy and safety of imputation methods for expanding CI datasets.

Main Methods:

• Utilized 7,451 audiograms from 32 institutions for CI candidacy evaluation.
• Assessed imputation models (univariate, interpolation, MICE, k-NN, gradient boosted trees, neural networks) using nested cross-validation.
• Defined a safe imputation threshold as root mean square error (RMSE) <10dB.

Main Results:

• Multiple Imputation by Chained Equations (MICE) consistently outperformed other models.
• MICE achieved an RMSE of 7.83 dB for imputing up to 6 missing features per audiogram.
• Imputation of 6 missing features, representing 99.3% of audiograms, increased dataset size 5.7-fold.

Conclusions:

• MICE provides safe and effective imputation for the majority of CI audiograms, well below the 10dB clinical threshold.
• Validated imputation models are essential for data-driven precision medicine in hearing healthcare.
• The findings suggest MICE is generalizable for future applications in large-scale audiometric data analysis.