Optimizing Clinical Trials with Digital Composite Covariates: A Case Study in Auditory Sciences

Background: New treatments are needed for individuals affected by hearing loss. Enhancing hearing loss phenotyping in clinical trials and increasing the diversity of patient populations will benefit programs focusing on screening, awareness, and education. Purpose: This study aimed to create a digital composite covariate (DCC) for pure-tone average (PTA) to enhance the screening process and prognostic capabilities for hearing loss clinical trials. Research Design and Study Sample: This study used cross-sectional data from the National Health and Nutrition Examination Survey. We downloaded, cleaned, and harmonized data across nine cohorts from 1999 to 2018 and identified 21,946 participants with audiometry data. Data Collection and Analysis: The DCC was created by predicting the PTA using simple-to-measure variables. A least absolute shrinkage and selection operator (LASSO) regression and an extreme gradient boosted (XGBoost) tree-based model were used. A stacking ensemble model was created by combining the predictions from LASSO and XGBoost. Results: The three machine learning models performed similarly. Stacking provided minute improvements over LASSO and XGBoost. The stacking ensemble model yielded a root mean square error of 7.63 dBHL (95 percent confidence interval [CI] 7.34, 7.92), a bias of 0.55 dB (95 percent CI 0.32, 0.76), and a correlation of 0.78 (95 percent CI 0.77, 0.80). A classification model was developed to predict the participants’ hearing status (i.e., no hearing loss vs. some hearing loss). Our model yielded an accuracy and balanced accuracy of 0.82 (95 percent CI 0.8, 0.83) and 0.84 (95 percent CI 0.83, 0.86), respectively, and an area under the receiving characteristic curve of 0.92 (95 percent CI 0.91, 0.93). Conclusions: The use of DCC prognostic factors for PTA can be used for phenotyping of hearing loss using simple-to-measure variables. Ultimately, this will aid in the optimization of clinical trials. Clinical Relevance Statement: This study has demonstrated that a person’s hearing ability can be predicted using easy-to-collect variables. In turn, we showed the potential to enhance the screening process for the recruitment of patients into audiology trials by identifying patient subpopulations that would benefit the most from the proposed treatment. In addition, the use of prognostic risk scores may contribute to sample size optimization, as well as the enhancement of auditory phenotyping and diversity of patient populations in clinical trials.