Parkinson disease subtypes based on short time series and multi-view clustering

KRALJEVSKA, MELANIJA (author), Robnik Šikonja, Marko (mentor), Valmarska, Anita (comentor)

Parkinson's disease (PD) is a progressive brain disorder which is characterized by movement problems such as tremor, stiffness, slowness of movement and dizziness, as well as non-motor symptoms, which include sleep disorders, constipation, problems concentrating, depression and emotional changes. Due to the clinical heterogeneity of PD, the existence of subtypes of PD patients has been addressed in many clinical and research studies and may contribute to a more personalized treatment and improved quality of life. We apply a methodology for discovering PD patient subtypes to patient data from the Fox Insight study (FI). The data sets are composed from questionnaires, containing patient symptoms and medication data collected through routine study visits. Dividing patients in subtypes can be translated to a problem of clustering time series data. We address this problem by using single-view clustering with k-means algorithm and multi-view spectral clustering. We describe the obtained subtypes with decision rules. Understanding decision making is crucial in medicine and we use decision trees as simple, explainable tools for describing subtypes. An important part of managing the disease is understanding the disease progression. By observing the patient's subtype changes between consecutive visits with skip-grams, we analyze the disease progression.

• Language: English
• Keywords: machine learning, data science, clustering, Parkinson's disease, multi-view learning, skip n-grams
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2021
• PID: 20.500.12556/RUL-125035
• COBISS.SI-ID: 53498371
• Publication date in RUL: 02.03.2021

Secondary language

• Language: Slovenian
• Title: Podtipi parkinsonove bolezni na podlagi kratkih časovnih vrst in gručenja z več pogledi

Abstract

Parkinsonova bolezen (PD) je progresivna možganska motnja, za katero so značilne motnje gibanja, kot so tremor, okorelost, počasnost in omotica, ter nemotorični simptomi, ki vključujejo motnje spanja, zaprtje, težave s koncentracijo, depresijo in čustvene spremembe. Zaradi klinične heterogenosti PD so v številnih kliničnih in raziskovalnih študijah obravnavali obstoj podtipov bolnikov s PD, kar lahko prispeva k bolj prilagojenemu zdravljenju in izboljšanju kakovosti življenja. Predstavljamo metodologijo za odkrivanje podtipov bolnikov s PD z uporabo podatkov o bolnikih iz študije Fox Insight (FI). Nabori podatkov izhajajo iz vprašalnikov, za katere z rutinskimi študijskimi obiski zbirajo podatki o bolnikovih simptomih in zdravilih. Razvrščanje pacientov v podtipe je v bistvu problem združevanja podatkov iz časovnih vrst. V naši nalogi problem rešujemo z algoritmom k-means in s spektralnim združevanjem v okviru učenja z več pogledi. Opis dobljenih podtipov dobimo z generiranjem pravil. Razumevanje odločanja je v medicini ključnega pomena, zato smo odločitvena drevesa uporabili kot preprosto, a razložljivo orodje za opis podtipov. Pomemben del obvladovanja bolezni je razumevanje napredovanja bolezni. Z opazovanjem prehodov pacientov med podtipi tekom zaporednih obiskov analiziramo napredovanje bolezni s pomočjo preskočnih n-gramov.

• Keywords: strojno učenje, podatkovne vede, razvrščanje, Parkinsonova bolezen, učenje z več pogledi, preskočni n-gram