Ekstrakcija medicinskega znanja iz tekstovnih opisov pri napovedovanju okužbe z rezistentno bakterijo

MIKUŠ, SANDI (author), Kukar, Matjaž (mentor), Beović, Bojana (comentor)

Problem odpornosti mikroorganizmov proti protimikrobnim sredstvom, je iz dneva v dan vse večji. Število mikroorganizmov, odpornih proti protimikrobnim sredstvom narašča hitreje od števila novih protimikrobnih sredstev. Če se pri bolniku izbere napačno protimikrobno sredstvo (zdravilo), se lahko odpornost mikroorganizmov še poveča. Mikroorganizmov odpornih proti protimikrobnim sredstvom je kar nekaj, vendar smo se v našem delu osredotočili izključno na primer Escherichie coli, ki izloča encime ESBL. Bolniki začnejo prejemati ustrezno protimikrobno sredstvo šele, ko prispejo mikrobiološki izvidi (po približno dveh dneh). Na podlagi zdravnikovega izvida v naravnem jeziku (sestavljenega iz anamneze in statusa pacienta) smo poskušali napovedati, ali je bolnik okužen s prej omenjeno bakterijo E. coli ali ne. Čeprav smo dosegli visoko specifičnost (90% za SVM oz. 86% za naivni Bayesov klasifikator), klasifikator zaradi prenizke senzitivnosti (28% za SVM oz. 33% za naivni Bayesov klasifikator) in občutljivosti problema ne more biti uporabljen. Problem vse večje odpornosti bakterij proti protimikrobnim sredstvom zahteva, da imamo visoko tako specifičnost kot tudi senzitivnost. Za izboljšanje modela bi bilo potrebno povečati število izvidov in po možnosti razširiti obseg podatkov z rezultati laboratorijskih preiskav.

• Language: Slovenian
• Keywords: Obdelava naravnega jezika, tekstovno rudarjenje, rezistentne bakterije, klasifikacija, strojno učenje, ESBL, Escherichia coli, zdravniški izvidi
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-100556
• Publication date in RUL: 27.03.2018

Secondary language

• Language: English
• Title: Extraction of medical knowledge from a full-text description for predicting resistant bacteria infection

Abstract

Resistant microorganisms are causing more and more problems in healthcare. Number of antibiotic-resistant microorganisms is growing faster than the number of newly discovered antibiotics. Wrongfully chosen antibiotics during treatment can also result in a greater resilience of the microorganisms. There exist several resistant microorganisms but we will focus on one, Escherichia coli which produces ESBL enzymes. Patients usually start receiving proper antibiotic treatment when doctors get microbiological reports (which takes around two days). We try to predict if a patient has the previously mentioned bacteria E. coli which produces ESBL enzymes, by using a medical report written in a natural language (which consists of the patient's history and status). Even if we achieved high specificity (90% with SVM and 86% with Naive Bayesian classifier) we can not use our models due to too low sensitivity (28% with SVM and 33% with Naive Bayesian classifier). Due to seriousness of the problem with resistant microorganisms it is required to have both metrics (specificity and sensitivity) high. In order to build better models we have to increase number of medical examination reports and maybe include additional results from other medical examinations.

• Keywords: Natural language processing, text-mining, resistant bacteria, classification, machine learning, ESBL, Escherichia coli, medical examination report