Differentiating patients with Parkinson’s disease and healthy individuals based on EEG analysis

Parkinsonova bolezen je kronična bolezen, za katero je značilno propadanje nevronov in posledičen primanjkljaj živčnega prenašalca dopamina. Parkinsonova bolezen lahko zelo oteži življenje bolnika, saj kot bolezen povzroča veliko različnih simptomov in težav, ki občutno poslabšajo kvaliteto življenja. Zato je ključnega pomena, da čim prej odkrijemo bolezen, jo začnemo tudi zdraviti in tako omilimo simptome. Kot ena izmed potencialnih diagnostičnih metod za Parkinsonovo bolezen se je v preteklosti pokazala analiza elektroencefalografskih (EEG) signalov. V okviru naše naloge smo poizkušali ločevati EEG posnetke zdravih oseb in oseb s Parkinsonovo boleznijo. V ta namen smo uporabili podatkovno bazo EEG posnetkov 25 oseb s Parkinsonovo boleznijo in 25 kontrolnih oseb. Problema smo se lotili tako, da smo uporabili signale s posameznih elektrod za posamezne posnetke, jih obdelali z različnimi metodami, vizualizirali in nato identificirali signale, kjer so razlike najbolj očitne. Sledila je analiza vseh posnetkov iz podatkovne baze, kjer smo naredili skupne pare korespondenčnih kontrolnih oseb in Parkinsonovih bolnikov, ki zdravil niso prejeli 15 ur, ter Parkinsonovih bolnikov, ki so zdravila jemali redno. V našem primeru so se kot najbolj informativni izkazali signali posneti na frontalnih elektrodah od F3 do F8. Pri klasifikaciji so bile najbolj uspešne nevronske mreže, kvadratna diskriminanta analiza in k-najbližjih sosedov. Izbrali smo celotno frekvenčno območje, ker smo tam dobili najboljše rezultate klasifikacije. Klasifikacijska točnost na testni množici je bila na nevronskih mrežah 93%, na kvadratni diskriminanti 89.5%, na klasifikatorju k-najbližjih sosedov pa 87.7%. Rezultati naše študije kažejo, da je ločevanje med osebami s Parkinsonovo boleznijo in zdravimi osebami na podlagi analize EEG posnetkov možno že v zgodnji fazi. Rezultati naše študije pa so tudi boljši od do sedaj objavljenih rezultatov klasifikacije na podatkovni množici, ki smo jo uporabili.Parkinson\u27s disease is a chronic disease characterised by the deterioration of neurons and a consequent deficiency of the neurotransmitter dopamine. Parkinson\u27s disease can make patient\u27s life very difficult, since it causes a wide range of symptoms and problems that significantly impair quality of life. It is therefore crucial to detect and treat the disease as early as possible to alleviate symptoms. In the past, the analysis of electroencephalographic (EEG) signals has emerged as one of the potential diagnostic methods for Parkinson\u27s disease. In our work, we tried to distinguish between EEG recordings of healthy subjects and subjects with Parkinson\u27s disease. For our study, we used a database of EEG recordings from 25 Parkinson\u27s patients and 25 control subjects. We approached the problem by using the signals from each electrode of the individual recordings, processing them with different methods, visualising them and then identified signals where the differences are the most obvious. We then analysed all the recordings from the database, creating pooled pairs of corresponding control subjects and Parkinson\u27s patients who had not received medication for 15 hours, and Parkinson\u27s patients who were taking medication regularly. In our case, the most informative signals were those recorded at frontal electrodes F3 to F8. Neural Networks, Quadratic Discriminant Analysis and k-Nearest Neighbors were the most successful for classification. We chose the whole frequency range because the best classification results were obtained there. The classification accuracy on the test set was 93% using the Neural Networks, 89.5% using the Quadratic Discriminant Analysis, and 87.7% using the k-Nearest Neighbors classifier. The results of our study suggest that discriminating between Parkinson\u27s disease and healthy subjects based on the analysis of EEG recordings is possible already at an early stage. The results of our study are also better than previously published classification results on the dataset we used