Investigation of machine learning methods for predicting surgical parameters in strabismus surgery

Abstract Purpose To minimize the variability of surgical outcomes in strabismus surgery we evaluated machine learning models that predict the dosage for surgical correction of horizontal non-paretic strabismus, based on preoperative features and tried to determine feature importance with explainable artificial intelligence. Methods First, a structured retrospective analysis of patients who had strabismus surgery between 2003 and 2022 at the university clinic of Ulm was performed. A final streamlined dataset of 767 patients was included in the second step. We built a multilayer perceptron using the functional class from torch neural network to evaluate the data and build model 0. The features were analyzed with approaches from explainable artificial intelligence to check the attributions for plausibility. Results The highest number of predictions in the acceptable range was achieved with the Broyden-Fletcher-Goldfarb-Shanno algorithm. For model 0 the root mean square error for both labels were: L1 = 0.54 mm and L2 = 0.71 mm. For L1 and L2, 69% and 55% of the data was predicted within the acceptable range respectively. The preoperative angle of deviation as the most important feature for dosages was confirmed by feature permutation as well as integrated gradients, followed by near angle of deviation, use of simultaneous cover test, Hirschberg test angle and refraction of own glasses used. Conclusion The developed neural network seems to offer a way to predict dosage in strabismus surgery and could assist surgeons in their decision making. A significant improvement of prediction accuracy is expected after increasing the data basis.