Wideband acoustic immittance (WAI) is an emerging test of middle-ear function with potential applications for neonates in screening and diagnostic settings. Previous large-scale diagnostic accuracy studies have assessed the performance of WAI against evoked otoacoustic emissions, but further research is needed using a more stringent reference standard. Research into suitable quantitative techniques to analyze the large volume of data produced by WAI is still in its infancy. Prediction models are an attractive method for analysis of multivariate data because they provide individualized probabilities that a subject has the condition. A clinically useful prediction model must accurately discriminate between normal and abnormal cases and be well calibrated (i.e., give accurate predictions). The present study aimed to develop a diagnostic prediction model for detecting conductive conditions in neonates using WAI. A stringent reference standard was created by combining results of high-frequency tympanometry and distortion product otoacoustic emissions.Design:
High-frequency tympanometry and distortion product otoacoustic emissions were performed on both ears of 629 healthy neonates to assess outer- and middle-ear function. Wideband absorbance and complex admittance (magnitude and phase) were measured at frequencies ranging from 226 to 8000 Hz in each neonate at ambient pressure using a click stimulus. Results from one ear of each neonate were used to develop the prediction model. WAI results were used as logistic regression predictors to model the probability that an ear had outer/middle-ear dysfunction. WAI variables were modeled both linearly and nonlinearly, to test whether allowing nonlinearity improved model fit and thus calibration. The best-fitting model was validated using the opposite ears and with bootstrap resampling.Results:
The best-fitting model used absorbance at 1000 and 2000 Hz, admittance magnitude at 1000 and 2000 Hz, and admittance phase at 1000 and 4000 Hz modeled as nonlinear variables. The model accurately discriminated between normal and abnormal ears, with an area under the receiver-operating characteristic curve (AUC) of 0.88. It effectively generalized to the opposite ears (AUC = 0.90) and with bootstrap resampling (AUC = 0.85). The model was well calibrated, with predicted probabilities aligning closely to observed results.Conclusions:
The developed prediction model accurately discriminated between normal and dysfunctional ears and was well calibrated. The model has potential applications in screening or diagnostic contexts. In a screening context, probabilities could be used to set a referral threshold that is intuitive, easy to apply, and sensitive to the costs associated with true- and false-positive referrals. In a clinical setting, using predicted probabilities in conjunction with graphical displays of WAI could be used for individualized diagnoses. Future research investigating the use of the model in diagnostic or screening settings is warranted.