Κατεύθυνση Βιοπληροφορική-Υπολογιστική ΒιολογίαLibrary of the School of Science
Καθηγήτρια Κωνσταντίνα Νικήτα
Τομέας Συστημάτων Μετάδοσης Πληροφορίας και Τεχνολογίας Υλικών ,
Τμήμα Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών , ΕθνικόΜετσόβιο Πολυτεχνείο
Ανάπτυξη ερμηνεύσιμων υπολογιστικών μοντέλων για την υποστήριξη κλινικών αποφάσεων στη διαχείριση του Σακχαρώδους Διαβήτη Τύπου 2 και της Περιφερικής Αρτηριακής Νόσου
Development of explainable computational models to support clinical decisions in the management of Type 2 Diabetes and Peripheral Arterial Disease
The present thesis aims at the development and evaluation of explainable clinical
decision support systems, based on artificial intelligence. Specifically, personalized
and interpretable models have been developed to assess the risk of Type 2 Diabetes
Mellitus (T.2.D.M) both short-term (3 years) and long-term (15 years) and the risk of
amputation after surgery in patients with Peripheral Artery Disease (P.A.D.). Both
T.2.D.M. and P.A.D. present high rates of prevalence and are often encountered as
comorbidities. The occurrence of T.2.D.M. can be prevented through the adoption of
effective behavioural lifestyle changes and body weight regulation, making thus the
detection of high-risk individuals of utmost importance. The management of P.A.D.
can be improved by providing to the health care professionals the risk of amputation
30 days after surgery empowering them in clinical decision making. that decisions can
be made in time.
The development and evaluation of T.2.D.M. risk prediction models has been
based on data from the public database of the Korean Genome and Epidemiology
Study. From a total of approximately 235,000 people who have participated in the
study, data corresponding to 1,000 randomly selected participants are available for
research purposes. The data has been collected by the members of the organization
from 2003 to 2017 with a frequency of follow up assessments up to 2 -4 years. The
data include demographic, clinical and lifestyle information. The short-term risk
prediction model takes into account the evolution of predisposing factors over the past
eight years and generates the possibility of disease onset within the next three years.
The long-term risk prediction model takes into account data at the baseline visit and
produces the probability of disease onset within the next 15 years. In order to achieve
personalization and transparency, the use of machine learning techniques (XGBoost,
LSTM) in combination with methods of interpreting the results extracted from the
models (LIME, SHAP Deep Explainer) have been investigated.
The model for assessing the risk of amputation in patients with P.A.D. has been
developed based on data granted from the Hippokration General Hospital of Athens. Data
include demographic, somatometric, clinical and lifestyle information from 73
patients. The development of the model has been based on the XGBoost method and
the interpretation of the results on the LIME technique. Ensemble learning techniques
have been applied to address the unbalanced nature of the datasets. The models have
been evaluated in terms of their discrimination ability and accurate production of the
explainable risk assessments.
Main subject category:
Type 2 Diabetes, Peripheral Arterial Disease, Cardiovascular Diseases, Explainable Artificial Intelligence, Risk Assessment, Decision Trees, Neural Networks, Ensemble Learning, XGBOOST, LSTM ,LIME,SHAP
Διπλωματική εργασία Ιωάννα Σταθάτου.pdf
File access is restricted until 2022-04-14.