Faculty of Computer Science and Engineering

Permanent URI for this communityhttps://repository.ukim.mk/handle/20.500.12188/5

The Faculty of Computer Science and Engineering (FCSE) within UKIM is the largest and most prestigious faculty in the field of computer science and technologies in Macedonia, and among the largest faculties in that field in the region. The FCSE teaching staff consists of 50 professors and 30 associates. These include many “best in field” personnel, such as the most referenced scientists in Macedonia and the most influential professors in the ICT industry in the Republic of Macedonia.

Browse

Settings

Has files: NoSubject: search.filters.subject.machine learning

Search Results

Now showing 1 - 8 of 8
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Trends from Minimally Invasive to Non-invasive Glucose Measurements
    (IEEE, 2020-09-28)
    ;
    Poposka, L.
    ;
    Guseva, E.
    ;
    ;
    Approximately 7% of the elderly over 45 are diagnosed in several diabetes forms and it is believed that 3% of the population is undiagnosed. As the number of diabetes diagnoses has been increasing, so too has the technology for managing the disease. The traditional way of measuring glucose levels is to apply a blood drop to a chemically treated, disposable ”test-strip” and insert it into an electronic device. In this study, we overview the latest technology for continuous glucose measurement, as they trend from minimally invasive to non-invasive glucose measurement methods.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Detecting Ventricular Beats with Machine Learning Models
    (IEEE, 2022-05-23)
    Tudjarski, Stojancho
    ;
    Stankovski, Aleksandar
    ;
    This paper aims at modeling a classifier of Ventricular heartbeats by experimenting with the most advanced classic binary classifiers in different scenarios for feature engineering. Methodology: The results were acquired based on experimenting with XGBoost and Random Forest algorithms, as two of the most advanced classifiers not based on neural networks. Although the annotated ECG data sets contain records with several heartbeat classes, we focus on a model that would distinguish V from others (Non-V heartbeats). Considering that we are dealing with a highly imbalanced data set, we applied the SMOTE algorithm for data enrichment to provide a better-balanced data set for training the model. To acquire better results, we added new calculated features, with and without feature selection. For feature selection, we used the Fisher Selector algorithm. Data: We used MIT-BIH Arrhythmia benchmark database, with train/test split according to the patient-oriented splitting approach that separates the original dataset into two subsets with approximately equal sizes and distribution of heartbeat types. Conclusion: The best results are achieved with XGBoost algorithm with original feature set. We achieved precision of 91.36%, recall of 88.31% and F1 score of 89.81%. Results showed that oversampling does not provide significantly better overall model performance. Still, we would recommend this approach since in practice, when dealing with imbalanced data sets, this leads to more robust models that perform better with data outside the training and test sets, such as when the model is used in production.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    From Continuous ECG Signals to Extracted Features for Machine Learning Models and Arrhythmia Annotations
    (IEEE, 2022-11-15)
    Tudjarski, Stojancho
    ;
    Stankovski, Aleksandar
    ;
    ;
    This paper describes the process of transforming an ECG signal as a continuous stream of numbers representing measured electrical voltages between the ECG electrodes into an output indicating the existence of arrhythmia. The ECG data stream is a structured array of converted analog signal values to digital data. Although this stream uses a continuous structure of numbers within a given range that depends on the bit resolution during conversion, it is still unstructured as a representation of the appearance of arrhythmia since it does not contain information about detected arrhythmia. This paper presents how to process ECG data, detect heartbeat annotations, and calculate various parameters for tabular-based data with a fixed number of columns to be used as input into ML-based algorithms. Our use case addresses an ML algorithm to detect atrial fibrillation arrhythmia, as an irregular heart rhythm. Practically a set of numbers in the ECG samples, which do not have structured arrhythmia annotations, is transformed into structured annotations. Experiments are conducted on the well-known ECG benchmark MIT-BIH Arrhythmia Database. The input data is resampled from 360 Hz to 125 Hz signal, and a signal processing algorithm is used to detect heartbeats, extracting a fixed set of features, and systematically forwarded to the feature selection ML methodologies to obtain atrial fibrillation annotations.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Improving Atrial Fibrillation Detection with Machine Learning Models
    (IEEE, 2025-06-02)
    Tudjarski, Stojancho
    ;
    ;
    Madevska Bogdanova, Ana
    ;
    Stankovski, Aleksandar
    In this paper, we focus on feature engineering to detect Atrial Fibrillation by determining whether the heart rhythm has irregularities without patterns. We experiment with a broad spectrum of features derived from the duration of heartbeat-to-heartbeat intervals in the benchmark electrocardiogram databases MIT-BIH Arrhythmia Database, MIT-BIH Atrial Fibrillation Database, and Long Term AF Database. The experiments included position-based features, fluctuation indices, standard deviations, mean average values, Shannon entropy, and statistical measures, such as compressed time series data length. The research questions are to detect the most influential features that result in the best-performing model and the impact of the training dataset. Our approach is to evaluate the model on a completely different dataset from the one it was trained on. Achieved F1 scores vary between 62,32% and 85.08%. The results prove that positional features increase the model's performance.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Overview of Interpretable and Explainable Artificial Intelligence for Atrial Fibrillation
    (IEEE, 2025-11-25)
    Tudjarski, Stojancho
    ;
    Angjelevska, Ana
    ;
    ;
    Madevska Bogdanova, Ana
    Accurate and interpretable detection of irregular work of the heart, such as atrial fibrillation, from electrocardiogram (ECG) signals is crucial for timely diagnosis and effective patient management. While machine learning (ML) models, particularly deep learning architectures, have achieved state-of-the-art performance in ECG arrhythmia classification, their black-box nature limits clinical adoption. This paper explores explainable artificial intelligence (XAI) techniques applicable to ML models trained on ECG data, highlighting both global and local interpretability approaches. We provide an overview of posthoc methods, including SHAP, LIME, PFI, and LIG, among others, treating various types of ECG recordings. As a practical case study, we present our findings analyzing the results of PFI and LIG methods applied to a transformer-based model fine-tuned for atrial fibrillation detection and explain its decision process. Our findings underscore the value of integrating XAI into ECG analysis pipelines to enhance transparency, foster clinician trust, and support more informed decision-making in cardiovascular diagnostics.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Bridging Gaps in Ligand Binding Affinity Prediction: Empirical Machine Learning Analyses
    (IEEE, 2025-06-02)
    Fetaji, Fjolla
    ;
    ;
    Predicting ligand binding affinity remains a critical challenge in computational drug discovery, as existing techniques often require extensive computational resources and are not readily generalizable to diverse protein-ligand systems. This study addresses three key gaps in current research: (1) the lack of versatile, data-efficient predictive models; (2) insufficient strategies for integrating protein and ligand structural information; and (3) limited methods for simultaneously improving accuracy and generalizability. By systematically reviewing recent advances in machine learning-including approaches derived from deep learning, graph-based methodologies, and hybrid frameworks-we show how emerging techniques enable higher predictive power and reduced computational cost. We leverage two large-scale, public datasets (PDBBind and BindingDB) to empirically evaluate a novel dual-model framework that integrates graph-based feature extraction and neural network regression. Comparative analyses illustrate how spatial and sequence representations contribute to model performance, achieving robust improvements in binding affinity prediction. The theory-based advantages of this approach demonstrate how it reveals both small-scale and wide-ranging relationships between proteins and their ligands and the operational benefits result in quicker medication development through decreased processing needs. The research confirms the necessity of building adaptable frameworks which unite structural data with sequence data for therapeutic advancement along clear research paths.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    RoBERTa for URL Classification: Enhancing Web Security and Content Filtering
    (IEEE, 2023-11-21)
    Ilievska, Joana
    ;
    Mihajloska Trpcheska Hristina
    ;
    Dobreva, Jovana
    ;
    ;
    The rising occurrence of malicious URLs on the internet is a major concern for users and their devices. To combat this problem, there is a need to develop effective methods to protect against these threats. The use of artificial intelligence (AI) presents an opportunity to leverage this technology to tackle this issue. This study proposes a URL classification model that uses RoBERTa transformer, an AI-based natural language processing technique, to classify URLs based on their intent. The model’s performance has been evaluated using various metrics, showcasing the potential of AI to assist in the creation of robust web security and online content filtering tools.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Advancing Image Spam Detection: Evaluating Machine Learning Models Through Comparative Analysis
    (MDPI AG, 2025-05-30)
    Jamil, Mahnoor
    ;
    ;
    ;
    Creutzburg, Reiner
    ;