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

Author: search.filters.author.Petrovski, Nikola

Search Results

Now showing 1 - 5 of 5
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Effect of Bézier Interpolation on Similarity of Heartbeats in Electrocardiograms
    (IEEE, 2022-11-15)
    ;
    Tonkovikj, Lucija
    ;
    Petrovski, Nikola
    Sampling frequency and bit resolution impact the presentation of heartbeats in electrocardiograms. One of the most challenging tasks in automated medical interpretation is the determination of the heartbeat type, whether the beat was initiated by a normal sequence, or by the ventricle. The usual way to detect the beat class is to analyze the morphological features including shape, width and height of the heartbeat. In this paper, we tackle this problem by checking the similarity of the heartbeats and set a research hypothesis that Bézier interpolation can improve the correctness of the similarity check. We found that the similarity check of two heartbeats can be improved a lot by using Bézier interpolation on the samples around the detected peak.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Scalability Evaluation of HPC Multi-GPU Training for ECG-based LLMs
    (IEEE, 2025-06-02)
    ;
    Petrovski, Nikola
    ;
    Training large language models requires extensive processing, made possible by many high-performance computing resources. This study compares multi-node and multi-GPU environments for training large language models using electrocardiogram data. It provides a detailed mapping of current frameworks for distributed deep learning in multi-node and multi-GPU settings, including Horovod from Uber, DeepSpeed from Microsoft, and the built-in distributed capabilities of PyTorch and TensorFlow. We compare various multi-GPU setups for different dataset configurations, utilizing multiple HPC nodes independently and focusing on scalability, speedup, efficiency, and overhead. The analysis leverages HPC infrastructure with SLURM, Apptainer (Singularity) containers, CUDA, PyTorch, and shell scripts to support training workflows and automation. We achieved a sub-linear speedup when scaling the number of GPUs, with values of 1.6x for two GPUs and 1.9x for four GPUs.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Deep Learning Image Embeddings for Ventricular Beat Classification
    (IEEE, 2025-11-25)
    Tudjarski, Stojancho
    ;
    Petrovski, Nikola
    ;
    ;
    Ventricular arrhythmias pose significant risks to cardiovascular health, necessitating accurate detection from electrocardiogram signals. This study investigates the effectiveness of vector embeddings derived from various convolutional neural networks and contrastive language-image pretraining-based models for classifying ventricular heartbeats presented as images. A comparative analysis revealed that the combination of the ResNet50 model and support vector machines achieved the top F1 score of 79.30%. The results provide insights into how deep neural network models can produce effective vector embeddings of heartbeats to train heartbeat classification models.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Optimizing Heartbeat Classification using Bézier Interpolation
    (IEEE, 2023-05-22)
    ;
    Petrovski, Nikola
    ;
    Tonkovikj, Lucija
    Analog to digital conversion of electrocardio-grams depends on the sampling frequency influencing the determination of a proper heartbeat location and precision of further digital processing. We set a research question to find the optimal number of interpolation points to reduce the mistakes in the similarity check of heartbeats and classify ventricular beats. In addition, another research question aims at finding the optimal number of interpolation points applying Bézier interpolation to reveal the optimal performance/cost ratio. Final research question is to find the sampling frequency that will reveal optimal performance in classification of ventricular beats. The experiments evaluate all neighbouring pairs of heartbeats from the standard benchmark MIT-BIH arrhythmia dataset resampled to a 125 Hz sampling frequency. The results show that even one more interpolation point, which corresponds to a sampling frequency of 250 Hz, will increase the performance versus the original 360 Hz sampling frequency. At the same time, the optimal is interpolation with additional five or seven points corresponding to 750 Hz, and 1000 Hz respectively. We found that a threshold value of 34 reveals the optimal performance to conclude a change between ventricular heartbeats and others, even in a 10-bit precision of the analog-digital conversion. The processing time and performance/cost-benefit analysis show that one interpolation point is the most beneficial.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    1D Convolutional Neural Network for Atrial Fibrillation Detection
    (IEEE, 2023-11-21)
    Petrovski, Nikola
    ;
    ;
    Tudzarski, Stojancho
    Biomedical Informatics faces a paramount challenge: efficient and accurate Atrial Fibrillation detection from a single-lead electrocardiogram. A cutting-edge Convolutional Neural Network model is constructed on differences of consecutive heart rate samples and extracting datasets with specific arrhythmia in the training process, being unique in the feature engineering and model development.Our model resulted in an F1 Score of 98.14%, evaluated by 10-fold cross-validation using the label evaluation method, comparable to other best results. However, we have proven that label-based methods are substantially higher than those based on inter-patient training/testing data split with different datasets, such as those used as ECG benchmarks detecting AFIB and the corresponding evaluation methods based on duration instead labels, such as those used in standards evaluating ambulatory ECG monitoring devices. Evaluation of our model with the inter-patient data split and duration-based evaluation method showed a 94.16% F1 score, which complies with standards and is high above the standard recommendations. This affirms the high efficacy and potential of the developed solution.