Repository of DUTH

Evaluation of ai-based diagnostic models for early detection of myelodysplastic syndromes and their correlation with laboratory-derived biomarkers (Master thesis)

Σμαραγδάκη, Νίκη/ Smaragdaki, Niki/ Βιο-ιατρικές και Μοριακές Επιστήμες στη Διάγνωση και Θεραπεία Ασθενειών

Myelodysplastic Syndromes (MDS) are a heterogeneous group of blood disorders characterized by abnormal cell production leading to peripheral blood cytopenias and an increased risk of progression to acute myeloid leukemia. Despite recent advances in understanding the molecular and cytogenetic aspects of the disease, currently, the only way to diagnose MDS is to use a combination of morphological, laboratory, and clinical criteria. Therefore, the search for additional biomarkers that could support the diagnostic process is particularly important. Recent studies suggest that oxidative stress and antioxidant imbalance may play an important role in the pathophysiology of MDS. The aim of this thesis is to investigate whether biomarkers related to oxidative stress and antioxidant defense, in combination with age, can help distinguish between patients with myelodysplastic syndromes and healthy individuals. The dataset included measurements of Reactive Oxygen Species (ROS), Malondialdehyde (MDA), Total Antioxidant Capacity (TAC) and Vitamin E, as well as the variable Age, from a total of 40 individuals (20 patients with MDS and 20 healthy controls). A hierarchical approach to analysis was adopted in the thesis. Initially, exploratory data analysis (EDA) was performed to assess the structure of the dataset, check the balance of the groups and search for possible correlations between the variables. Clustering was then performed to determine if there are clusters of data that point to different biological profiles. Then, supervised machine learning models (Artificial Neural Networks, Support Vector Machines, Random Forest and XGBoost) were trained and compared for binary classification of individuals into MDS group or control group. To ensure that the results of this work are valid and that there is no overfitting, internal validation methods and hyperparameter optimization were applied. Also, the influence of each factor on the final diagnostic result was studied, aiming to identify the biomarkers whose contribution was the most important. In this way, one can identify the effectiveness of the algorithms and evaluate the role of individual factors in the process. The results showed that the MDA index is a very strong discriminative biomarker. In addition, the indicators related to oxidative stress, and especially the levels of ROS and Vitamin E, presented the most pronounced differences between the two groups and the greatest contribution to the classification models. The machine learning models showed very good diagnostic performance, with AUC values up to 0.995 and accuracy exceeding 90% in cross-validation procedures. The findings of the diploma thesis show that oxidative stress markers, especially when combined with machine learning methods, can contribute substantially to the discrimination between MDS patients and healthy individuals. These results are encouraging for further exploration of multivariate approaches based on biomarkers as complementary diagnostic tools.