Authors: Apurv Shukla¹, Amy Paolleti², Pravansu Mohanty³, Devin McCaslin²
¹University of Michigan; ²University of Michigan Medicine; ³University of Michigan, Dearborn
Background: Advances in computational audiology rely on real-world clinical data to develop predictive models for cochlear implant (CI) care. However, existing clinical registries are often poorly aligned with these goals. This study analyzes a 25-year registry of 2,039 adult CI patients to characterize the documentation gaps—such as etiology, imaging, and longitudinal speech outcomes—that limit downstream modeling.
Method: We evaluated the structure and completeness of a single-center registry, identifying systematic measurement bias and non-random missingness patterns. We analyzed how variables like etiology stratify patients and investigated the frequency of structured documentation for key clinical indicators, comparing current records against the requirements for effective machine learning and decision-support pipelines.
Results: Critical variables are sparsely recorded: etiology is undocumented in 75% of patients, imaging findings are structured in only 13%, and longitudinal speech outcomes are available for fewer than 5%. These patterns reflect clinical workflows and historical practices rather than random omissions. Naive analyses of such data risk biased inference, making standard machine learning generalization unlikely.
Conclusion: We argue for a “measurement-aware” registry design paradigm where data collection is explicitly aligned with modeling objectives. A concrete roadmap for future CI registries must include structured etiology capture, standardized imaging annotations, and consistent longitudinal speech outcome tracking to support personalized intervention and robust predictive modeling.
