Determinación mediante minería de datos del núcleo subtalámico utilizando registros MER de cirugías para la implantación de neuroestimuladores en pacientes con Parkinson

A deep brain stimulator (DBS) is an implantable neuro stimulator for the treatment of motor disorders of Parkinson's disease (PD). Using minimally invasive stereotactic surgery techniques, two stimulation electrodes are usually implanted in the Subthalamic nuclei (STN) of both cerebral hemispheres. The detection of STN is a complex and crucial task for the success of the therapy with a DBS, which is carried out in most institutions through the analysis of the patterns of signals acquired using microelectrodes register (MER). The hypothesis of the present thesis is defined like it is possible to obtain supervised classification models with a good performance, making use of the ensemble classifiers and an adequate selection of characteristics, in order that they can be used as a support tool to detect the MER signals corresponding to the STN during surgery to implant a DBS. MER signals were acquired from 22 patients with PD who underwent bilateral surgery for the implantation of a DBS in STN. As the MER descended towards the STN, whose initial coordinates were established by magnetic resonance imaging, two specialists established whether the records observed at each depth level corresponded to 'STN' or 'non-STN'. After the surgery, each record was divided into windows of 1 s overlapped to 50%, which were calculated and standardized off line 16 temporal characteristics. With the database obtained, the classifiers were trained with the data of 21 patients (training data) leaving the patient's data excluded for validation (validation data). This process was repeated 22 times leaving a different patient at a time in each training-validation set. The trained stacking algorithm with the 16 temporal characteristics obtained the best results using 3 different classifiers as a base: K- nearest neighbors, CART and bagging decision trees and random forest as meta classifier. This stacking classifier obtained an average accuracy (ACC) of 94.6%, a specificity (ESP) of 95.6% and a sensitivity (SEN) of 95.6%. The branch and bound feature selection algorithm was the best for the thesis data and allowed to select 6 characteristics of the total 16 that improved the stacking performance, which one obtained an ACC of 95%, ESP of 96% and SEN of 94%. A combined model of supervised classification was obtained using the stacking and branch and bound methodology, whose performance and validation times indicate that it can be used, with excellent performance, in an automatic classification process to detect STN signals from the temporal characteristics of cerebral electrical signals of the MER.