Exploring the Potential of Machine Learning for the Diagnosis of Balance Disorders Based on Centre of Pressure Analyses

Abstract

Balance disorders are caused by several factors related to functionality deficits in one or multiple sensory systems such as vision, vestibular, and somatosensory systems. Patients usually have difficulty explaining their dizziness, often using ambiguous words to describe their symptoms. A common practice by clinicians is to objectively evaluate the patient’s dizziness by applying the Sensory Organization Test (SOT), which measures the contribution of each sensory system (vestibular, visual, somatosensory). The SOT protocol can record up to 2000 measurements in 20 s to generate the Equilibrium Score (EQS) with its five load sensors. EQS is an indicator that reflects how well a patient can maintain balance. However, its calculation only considers two instances from these 2000 measurements that reflect the maximum anterior and posterior sway angle during the test performance; therefore, there is an opportunity to perform further analysis. This article aims to use the Centre of Pressure (COP) time series generated by the SOT and describes a methodology to pre-process and reduce the dimensionality of this raw data and use it as an input for machine learning algorithms to diagnose patients with balance disorder impairments. After applying this methodology to data from 475 patients, the logistic regression model (LR) produced the highest f1-score with 76.47%, and the support vector machine (SVM) performed almost as well, with an f1-score of 76.19%.

Keywords

• balance disorder
• approximate entropy
• empirical mode decomposition
• machine learning model