Objective Assessment of Pathological Voice Using Artificial Intelligence Based on the GRBAS Scale.

OBJECTIVES

The validity and reliability of the psychological assessment of auditory perceptions, as typified by the grade, roughness, breathiness, asthenia, and strain (GRBAS) scale, have been widely recognized. However, due to their subjective nature, inter- and intra-examiner reliability are unavoidable. In this study, we aimed to add objectivity to the GRBAS scale using artificial intelligence and to compare the accuracy of two methods-one based on Google's TensorFlow and another based on Apple's Core ML.

METHODS

The GRBAS scale of 1,377 vowel samples was evaluated and used as training data to create a machine learning model. We used TensorFlow and Apple's Create ML to create two machine learning models and examined the difference in their accuracies for classifying the severity of pathological Voice data based on the GRBAS scale.

RESULTS

Absolute comparisons are difficult to make because of the difference in methods; however, both training models could objectively evaluate GRBAS scales and were statistically correlated in G and B.

CONCLUSION

While TensorFlow requires creation of a training model from scratch, Create ML is a relatively easy way to create a training model for voice by adding training data for GRBAS scales to an existing training model for sounds. Although the data handling and learning methods are different, both models performed well. Findings from this study could be used for medical screening purposes, and there is the potential to change the clinical approach to voice diagnostics in the future.