A comparison of classifiers in biomedical signal processing as a decision support system in disc hernia diagnosis.

The aim of this research was to investigate the best methodology for disc hernia diagnosis using foot force measurements from the designed platform. Based on the subjective neurological examination that examines muscle weakness on the nerve endings of the skin area on feet and concludes about origins of nerve roots between spine discs, a platform for objective recordings of the aforementioned muscle weakness has been designed. The dataset included 33 patients with pre-diagnosed L4/L5 and L5/S1 disc hernia on the left or the right side, confirmed with the MRI scanning and neurological exam. We have implemented 5 different classifiers that were found to be the most suitable for smaller dataset and investigated the accuracy of classification depending on the normalization method, linearity/non-linearity of the algorithm, and dataset splitting variation (32-1, 31-2, 30-3, 29-4 patients for training and testing, respectively). The classifier is able to distinguish between four different diagnoses L4/L5 on the left side, L4/L5 on the right side, L5/S1 on the left side and L5/S1 on the right side, as well as to recognize healthy subjects (without disc herniation). The results show that non-linear algorithms achieved better accuracy in comparison to tested linear classifiers, suggesting the expected non-linear connection between the foot force values and the level of disc herniation. Two algorithms with highest accuracy turned out to be Decision Tree and Naïve Bayes, depending on the normalization method. The system is also able to record and recognize improvements in muscle weakness after surgical operation and physical therapy.