Predictive Modeling of DWT-decomposed ALS-EMG Features Using Group Method of Data Handling

Identification of Amyotrophic Lateral Sclerosis disorder using electromyography signal data requires accurate classification models. Traditional approaches often necessitate manual model control and do not provide the designer with a set of practical classification equations. In this paper, a Group Method of Data Handling-based methodology is proposed which automatically builds the model by heuristically estimating optimum quadratic polynomial equations from available empirical data. Based on guidelines in literature and diligent trial and error assessment, a reduced set of most favorable features was tuned and assembled after extraction from a selected sub-band of Discrete Wavelet Transform decompositions. The generated model equations are an approximate representation of relationship between training data and corresponding class labels. For evaluation of prediction performance, computational experiments were conducted on test data and results were compiled on the grounds of Leave-one-out cross-validation criterion. The trained model effectively identified diseased and healthy data groups with Sensitivity, Specificity and Accuracy of 90.34%, 94.89% and 92.55% which is an improvement of 0.68%, 2.92% and 1.77% respectively, over the results of a conventional Multilayer Perceptron classifier tested under similar experimental conditions. These outcomes indicate that the proposed approach could be extended to real clinical setting to assist the clinicians in making precise diagnosis and potentially compete with leading automated diagnostic support systems.