Multi-class Classification and Feature Analysis of Symptoms in a Digital Assessment of Tremor

The first multi-class approach to evaluation of tremor severity based on data remotely recorded by a mobile- or tablet-based drawing application is developed which seeks to differentiate between healthy subjects, essential tremor patients receiving deep brain stimulation treatment, and those same patients with treatment disabled. Linear discriminant analysis and gradient boosting classification algorithms are implemented on the features extracted from the sub-bands of Discrete Wavelet Transform of the signals, with overall accuracy of 97.04%. This highly effective classification algorithm erroneously classified ”treated” samples as ”healthy”, implying that DBS treatment is capable of completely eliminating tremor.