Unsupervised Machine Learning in Classification of Neurobiological Data

Abstract

In many cases of neurophysiological data analysis, the best results can be obtained using supervised machine learning approaches. Such very good results were obtained in detection of neurophysiological recordings recorded within Subthalamic Nucleus (${ STN}$) during deep brain stimulation (DBS) surgery for Parkinson disease. Supervised machine learning methods relay however on external knowledge provided by an expert. This becomes increasingly difficult if the subject’s domain is highly specialized as is the case in neurosurgery. The proper computation of features that are to be used for classification without good domain knowledge can be difficult and their proper construction heavily influences quality of the final classification. In such case one might wonder whether, how much and to what extent the unsupervised methods might become useful. Good result of unsupervised approach would indicate presence of a natural grouping within recordings and would also be a further confirmation that features selected for classification and clustering provide good basis for discrimination of recordings recorded within Subthalamic Nucleus (${ STN}$). For this test, the set of over 12 thousand of brain neurophysiological recordings with precalculated attributes were used. This paper shows comparison of results obtained from supervised - random forest based - method with those obtained from unsupervised approaches, namely K-Means and Hierarchical clustering approaches. It is also shown, how inclusion of certain types of attributes influences the clustering based results.