Speech Analysis as a Decision Support System in Healthcare for Detecting Mild Traumatic Brain Injury

Abstract

Mild Traumatic Brain Injury (mTBI), also known as concussion, is a prevalent neurological condition with significant public health consequences. Accurate and timely diagnosis of mTBI is crucial for effective management and prevention of long-term complications. However, current diagnostic methods, including clinical assessments and neuroimaging techniques, have limitations in terms of subjectivity, sensitivity, and accessibility. This thesis explores the potential of speech analysis as a non-invasive and accessible tool for mTBI diagnosis. The feature selection process highlighted the importance of specific speech tasks and acoustic features in concussion detection, particularly those related to timing variability, prosody, pitch variation, intensity deviation, and speech motor control. The findings of this research suggest that subtle changes in speech patterns, often undetectable to human ear, can be detected and analysed using Machine Learning algorithms to aid in concussion diagnosis. By analysing selected acoustic features, two Machine Learning models, Support Vector Machine (SVM) and Multilayer Perceptron (MLP), were developed to differentiate between concussed and non-concussed individuals. Although the MLP model slightly outperformed the SVM model in terms of accuracy (81.59% vs. 80.75%), the SVM model's marginally higher AUC-ROC value (89.77% vs. 88.45%) suggests a potentially better overall performance due to its ability to better distinguish between the two classes across various classification thresholds. This study represents a significant step towards developing a more accurate, efficient, and accessible diagnostic tool for mTBI, with the potential to improve patient outcomes and reduce the long-term consequences of this injury.