Risk stratification of patients for development of cardiac diseases using knowledge transfer

Papič, Aleš (author), Bosnić, Zoran (mentor)

This thesis addresses the risk stratification of patients for development of cardiac diseases using machine learning methods. Our approaches modify existing methodologies, such as semi-supervised learning, active learning, fuzzy learning and supervised clustering. Using these methods we perform knowledge transfer on partially labeled data. We use the posterior class probability and local modeling of prediction error to strategically select training examples. Evaluation is performed on public heart disease data set and on data from peripheral arterial disease survival study. During the evaluation process, we use different ratios of labeled examples. The results show that our approaches increase the inductive performance compared to learning algorithms trained exclusively on labeled data.

• Language: English
• Keywords: risk stratification, cardiovascular diseases, machine learning, knowledge transfer, semi-supervised learning, active learning, fuzzy learning, supervised clustering, partially labeled examples
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2019
• PID: 20.500.12556/RUL-110024
• COBISS.SI-ID: 1538330051
• Publication date in RUL: 11.09.2019

Secondary language

• Language: Slovenian
• Title: Napovedovanje tveganja pacientov za razvoj srčnih obolenj s prenosom znanja

Abstract

V delu obravnavamo stratifikacijo tveganja za razvoj srčnih bolezni s pomočjo metod strojnega učenja. Naši pristopi uporabijo obstoječe metodologije, kot so delno-nadzorovano učenje, aktivno učenje, mehko učenje in nadzorovano gručenje. Pristopi izvajajo prenos znanja na delno označenih podatkih z izborom učnih primerov. Za strateško izbiro primerov uporabimo aposteriorno verjetnost razreda in oceno lokalne napake napovedi. Pristope analiziramo na javnih podatkih in podatkih s študije preživetja obolelih z periferno arterijsko boleznijo. Med eksperimentalno analizo uporabljamo različna razmerja označenih in neoznačenih primerov. Rezultati kažejo, da naši pristopi izboljšajo induktivno točnost v primerjavi z učnimi algoritmi, ki so učeni izključno z uporabo označenih primerov.

• Keywords: stratifikacija tveganja, bolezni srca in ožilja, strojno učenje, prenos znanja, delno-nadzorovano učenje, aktivno učenje, mehko učenje, nadzorovano gručenje, delno označeni primeri