Machine learning method for prediction of hearing improvement after stapedotomy

Introduction: Stapedotomy is considered for patients with conductive hearing loss of at least 15 dB in frequencies 250 to 1000 Hz or higher. In some cases, it is still difficult to decide whether surgery (stapedotomy) should be performed.

Methods: A total of 123 ear measurements from 79 female (approximately 64 %) and 44 male (approximately 36 %) subjects were included in the final dataset. We developed a machine learning method that predicts a patient"s post-operative hearing quality after stapedotomy, based on their pre-operative hearing quality and other features. Feature construction and feature selection methods were used extensively. Linear regression, ridge regression, lasso regression, random forest and gradient boosting models were applied. For better prediction, a separate regressor was trained to predict each postoperative hearing intensity component.

Results: The most successful predictions were made at air conduction frequencies between 1000 and 3000 Hz, with mean absolute errors of around 5 dB and in the case of ridge regression Adjusted - R2 between 0.7 and 0.8. On average, the ridge regression and random forest methods achieved the highest prediction accuracy. Comparing the cross validations (CV) and test performance sections of the table, it can be seen that the CV scores are on average better than the test scores. The average CV mean absolute error (MAE) across all classes is 6.24, whereas the average test MAE is 7.18.

Discussion: The proposed system shows a potential for support in recommending stapedotomy procedures. Overall, the post-operative bone-conduction hearing intensities proved to be challenging to predict. In its current state, this system can be adapted and trained on surgical data from a single surgeon.

Nein.