Objectives:During an initial diagnostic assessment of an ear with normal otoscopic exam, it can be difficult to determine the specific pathology if there is a mechanical lesion. The audiogram can inform of a conductive hearing loss but not the underlying cause. For example, audiograms can be similar between the inner-ear condition superior canal dehiscence (SCD) and the middle-ear lesion stapes fixation (SF), despite differences in pathologies and sites of lesion. To gain mechanical information, wideband tympanometry (WBT) can be easily performed noninvasively. Absorbance, the most common WBT metric, is related to the absorbed sound energy and can provide information about specific mechanical pathologies. However, absorbance measurements are challenging to analyze and interpret. This study develops a prototype classification method to automate diagnostic estimates. Three predictive models are considered: one to identify ears with SCD versus SF, another to identify SCD versus normal, and finally, a three-way classification model to differentiate among SCD, SF, and normal ears.Design:Absorbance was measured in ears with SCD and SF as well as normal ears at both tympanometric peak pressure (TPP) and 0 daPa. Characteristic impedance was estimated by two methods: the conventional method (based on a constant ear-canal area) and the surge method, which estimates ear-canal area acoustically.Classification models using multivariate logistic regression predicted the probability of each condition. To quantify expected performance, the condition with the highest probability was selected as the likely diagnosis. Model features included: absorbance-only, air-bone gap (ABG)-only, and absorbance+ABG. Absorbance was transformed into principal components of absorbance to reduce the dimensionality of the data and avoid collinearity. To minimize overfitting, regularization, controlled by a parameter lambda, was introduced into the regression. Average ABG across multiple frequencies was a single feature.Model performance was optimized by adjusting the number of principal components, the magnitude of lambda, and the frequencies included in the ABG average. Finally, model performances using absorbance at TPP versus 0 daPa, and using the surge method versus constant ear-canal area were compared. To estimate model performance on a population unknown by the model, the regression model was repeatedly trained on 70% of the data and validated on the remaining 30%. Cross-validation with randomized training/validation splits was repeated 1000 times.Design:Absorbance was measured in ears with SCD and SF as well as normal ears at both tympanometric peak pressure (TPP) and 0 daPa. Characteristic impedance was estimated by two methods: the conventional method (based on a constant ear-canal area) and the surge method, which estimates ear-canal area acoustically.Classification models using multivariate logistic regression predicted the probability of each condition. To quantify expected performance, the condition with the highest probability was selected as the likely diagnosis. Model features included: absorbance-only, air-bone gap (ABG)-only, and absorbance+ABG. Absorbance was transformed into principal components of absorbance to reduce the dimensionality of the data and avoid collinearity. To minimize overfitting, regularization, controlled by a parameter lambda, was introduced into the regression. Average ABG across multiple frequencies was a single feature. Model performance was optimized by adjusting the number of principal components, the magnitude of lambda, and the frequencies included in the ABG average. Finally, model performances using absorbance at TPP versus 0 daPa, and using the surge method versus constant ear-canal area were compared. To estimate model performance on a population unknown by the model, the regression model was repeatedly trained on 70% of the data and validated on the remaining 30%. Cross-validation with randomized training/validation splits was repeated 1000 times.Design:Absorbance was measured in ears with SCD and SF as well as normal ears at both tympanometric peak pressure (TPP) and 0 daPa. Characteristic impedance was estimated by two methods: the conventional method (based on a constant ear-canal area) and the surge method, which estimates ear-canal area acoustically.Classification models using multivariate logistic regression predicted the probability of each condition. To quantify expected performance, the condition with the highest probability was selected as the likely diagnosis. Model features included: absorbance-only, air-bone gap (ABG)-only, and absorbance+ABG. Absorbance was transformed into principal components of absorbance to reduce the dimensionality of the data and avoid collinearity. To minimize overfitting, regularization, controlled by a parameter lambda, was introduced into the regression. Average ABG across multiple frequencies was a single feature.Model performance was optimized by adjusting the number of principal components, the magnitude of lambda, and the frequencies included in the ABG average. Finally, model performances using absorbance at TPP versus 0 daPa, and using the surge method versus constant ear-canal area were compared. To estimate model performance on a population unknown by the model, the regression model was repeatedly trained on 70% of the data and validated on the remaining 30%. Cross-validation with randomized training/validation splits was repeated 1000 times.Results:The model differentiating between SCD and SF based on absorbance-only feature resulted in sensitivities of 77% for SCD and 82% for SF. Combining absorbance+ABG improved sensitivities to 96% and 97%. Differentiating between SCD and normal using absorbance-only provided SCD sensitivity of 40%, which improved to 89% by absorbance+ABG. A three-way model using absorbance-only correctly classified 31% of SCD, 20% of SF and 81% of normal ears. Absorbance+ABG improved sensitivities to 82% for SCD, 97% for SF and 98% for normal. In general, classification performance was better using absorbance at TPP than at 0 daPa.Conclusion:The combination of wideband absorbance and ABG as features for a multivariate logistic regression model can provide good diagnostic estimates for mechanical ear pathologies at initial assessment. Such diagnostic automation can enable faster workup and increase efficiency of resources.
