Stojkoska, Biljana

Preferred name

Stojkoska, Biljana

Official Name

Stojkoska, Biljana

Main Affiliation

biljana.stojkoska@finki.ukim.mk

Now showing 1 - 10 of 21

Some of the metrics are blocked by your
consent settings
A study on appropriate segment length for generalized cuff-less blood pressure estimation from ECG features
(IEEE, 2024-05-20)
Kuzmanov, Ivan
;
;
Lamenski, Petre
;
;
Madevska Bogdanova, Ana
Blood pressure (BP) refers to the pressure exerted on the blood vessels as blood travels through the body. Our ultimate goal is to build a stable model for BP estimation as part of a triage process. In this study, we experiment to determine a suitable signal segment only from electrocardiogram (ECG) signals, to ensure a fast and reliable process of the BP estimation. The used dataset contains only high-quality ECG and arterial blood pressure (ABP) signals extracted from the Medical Information Mart for Intensive Care, MIMIC II and MIMIC III databases by our methodology. It was processed three times using similar machine learning (ML) methodologies, with different segment lengths. Three different datasets are generated using a non-overlapping window with a size of 8, 15, and 30 seconds, with the same ECG features. Several linear and nonlinear Machine Learning models are built on these datasets, and their results are compared. Our best results were obtained by a light gradient-boosting machine (LightGBM) regression model trained on the 30-second dataset. The model achieves Mean Absolute Error (MAE) of 10.87, 6.55, and 7.29, and Root Mean Squared Error (RMSE) of 14.49, 8.68, and 9.68 for systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean arterial pressure (MAP), respectively. The results of our experiment indicate that a duration of 30 seconds is the minimum length that provides informative features, fulfilling the need for real-time delivery.
Some of the metrics are blocked by your
consent settings
Forecasting air pollution with deep learning with a focus on impact of urban traffic on PM10 and noise pollution
(Public Library of Science (PLoS), 2024)
Kostadinov, Martin
;
;
;
Coelho, Paulo Jorge
;
Air pollution constitutes a significant worldwide environmental challenge, presenting threats to both our well-being and the purity of our food supply. This study suggests employing Recurrent Neural Network (RNN) models featuring Long Short-Term Memory (LSTM) units for forecasting PM10 particle levels in multiple locations in Skopje simultaneously over a time span of 1, 6, 12, and 24 hours. Historical air quality measurement data were gathered from various local sensors positioned at different sites in Skopje, along with data on meteorological conditions from publicly available APIs. Various implementations and hyperparameters of several deep learning models were compared. Additionally, an analysis was conducted to assess the influence of urban traffic on air and noise pollution, leveraging the COVID-19 lockdown periods when traffic was virtually non-existent. The outcomes suggest that the proposed models can effectively predict air pollution. From the urban traffic perspective, the findings indicate that car traffic is not the major contributing factor to air pollution.
Some of the metrics are blocked by your
consent settings
Forecasting air pollution with deep learning with a focus on impact of urban traffic on PM10 and noise pollution
(Public Library of Science, 2024-12-10)
Kostadinov, Martin
;
;
;
Coelho, Paulo Jorge
;
Air pollution constitutes a significant worldwide environmental challenge, presenting threats to both our well-being and the purity of our food supply. This study suggests employing Recurrent Neural Network (RNN) models featuring Long Short-Term Memory (LSTM) units for forecasting PM10 particle levels in multiple locations in Skopje simultaneously over a time span of 1, 6, 12, and 24 hours. Historical air quality measurement data were gathered from various local sensors positioned at different sites in Skopje, along with data on meteorological conditions from publicly available APIs. Various implementations and hyperparameters of several deep learning models were compared. Additionally, an analysis was conducted to assess the influence of urban traffic on air and noise pollution, leveraging the COVID-19 lockdown periods when traffic was virtually non-existent. The outcomes suggest that the proposed models can effectively predict air pollution. From the urban traffic perspective, the findings indicate that car traffic is not the major contributing factor to air pollution.
Some of the metrics are blocked by your
consent settings
A study on appropriate segment length for generalized cuff-less blood pressure estimation from ECG features
(IEEE, 2024-05-20)
Kuzmanov, Ivan
;
;
Lamenski, Petre
;
;
Madevska Bogdanova, Ana
Blood pressure (BP) refers to the pressure exerted on the blood vessels as blood travels through the body. Our ultimate goal is to build a stable model for BP estimation as part of a triage process. In this study, we experiment to determine a suitable signal segment only from electrocardiogram (ECG) signals, to ensure a fast and reliable process of the BP estimation. The used dataset contains only high-quality ECG and arterial blood pressure (ABP) signals extracted from the Medical Information Mart for Intensive Care, MIMIC II and MIMIC III databases by our methodology. It was processed three times using similar machine learning (ML) methodologies, with different segment lengths. Three different datasets are generated using a non-overlapping window with a size of 8, 15, and 30 seconds, with the same ECG features. Several linear and nonlinear Machine Learning models are built on these datasets, and their results are compared. Our best results were obtained by a light gradient-boosting machine (LightGBM) regression model trained on the 30-second dataset. The model achieves Mean Absolute Error (MAE) of 10.87, 6.55, and 7.29, and Root Mean Squared Error (RMSE) of 14.49, 8.68, and 9.68 for systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean arterial pressure (MAP), respectively. The results of our experiment indicate that a duration of 30 seconds is the minimum length that provides informative features, fulfilling the need for real-time delivery.
Some of the metrics are blocked by your
consent settings
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.
Some of the metrics are blocked by your
consent settings
Deep Dive into Invoice Intelligence: A Benchmark Study of Leading Models
(Springer Nature, 2024-08-01)
Bajrami, Merxhan
;
;
;
Invoice data extraction is crucial in contemporary business environments, streamlining the processing and organization of voluminous invoice data for effi-cient management and decision-making. This paper presents a thorough benchmark study, evaluating the performance of five state-of-the-art models-LayoutLM, LiLT, Donut, Yolov8x, and Yolov5x—in the automated extraction of data from invoices. Leveraging transfer learning, each model is fine-tuned and assessed based on a private dataset comprising diverse invoice types. The evaluation metrics encompass precision, recall, F1-score, and accuracy, providing a comprehensive performance analysis. Results indicate that Yolov8x outperforms its counterparts with a com-mendable accuracy of 0.913 and an F1-score of 0.928, whereas LiLT trails with an accuracy of 0.537. Furthermore, the paper sheds light on the real-world deployment performance of these models, considering factors such as GPU usage, inference time on GPU and CPU, loading time, and loading GPU usage. LayoutLM demonstrates balanced performance with moderate resource consumption, while other models exhibit varied efficiencies and resource utilizations. Through this investigation, the paper furnishes invaluable insights into the strengths and limitations of each model, offering guidance for practitioners in selecting the optimal model tailored to specific invoice data extraction tasks and operational constraints.
Some of the metrics are blocked by your
consent settings
Forecasting air pollution with deep learning with a focus on impact of urban traffic on PM10 and noise pollution
(Public Library of Science (PLoS), 2024)
Kostadinov, Martin
;
;
;
Coelho, Paulo Jorge
;
Air pollution constitutes a significant worldwide environmental challenge, presenting threats to both our well-being and the purity of our food supply. This study suggests employing Recurrent Neural Network (RNN) models featuring Long Short-Term Memory (LSTM) units for forecasting PM10 particle levels in multiple locations in Skopje simultaneously over a time span of 1, 6, 12, and 24 hours. Historical air quality measurement data were gathered from various local sensors positioned at different sites in Skopje, along with data on meteorological conditions from publicly available APIs. Various implementations and hyperparameters of several deep learning models were compared. Additionally, an analysis was conducted to assess the influence of urban traffic on air and noise pollution, leveraging the COVID-19 lockdown periods when traffic was virtually non-existent. The outcomes suggest that the proposed models can effectively predict air pollution. From the urban traffic perspective, the findings indicate that car traffic is not the major contributing factor to air pollution.
Some of the metrics are blocked by your
consent settings
Trilateration and Multilateration Based Localization of Wireless Devices
(Ss Cyril and Methodius University in Skopje, Faculty of Computer Science and Engineering, Republic of North Macedonia, 2023-07)
Rusev, Nikola
;
Many new techniques for localization of wireless sensor networks have been developed in the past few decades, the most popular of which is multilateration – a simple technique that determines the position of one node or sensor in space using data collected from other nodes or sensors in the same network. In this paper, we propose a new localization algorithm based on multilateration, named as ‘Moving Point Algorithm’, which logarithmically searches the space to determine the location of a given node in the sensor network. The simulation of out algorithm shows that it outperforms traditional multilateration by means of localization accuracy and percentage of localized nodes.
Some of the metrics are blocked by your
consent settings
A Comprehensive Analysis of LayoutLM and Donut for Document Classification
(Ss Cyril and Methodius University in Skopje, Faculty of Computer Science and Engineering, Republic of North Macedonia, 2023-07)
Bajrami, Merxhan
;
;
;
Document classification is important in everyday life as it allows for efficient management and organization of vast amounts of digital documents, saving time and resources. This task is essential for businesses, organizations, and individ uals who handle large volumes of data and need to quickly retrieve and analyze specific information. AI-based document classification can help organizations better manage and organize their digital assets, improve information retrieval, and make better business decisions based on the insights derived from the classified documents. This paper compares the performance of two transformer-based models, LayoutLM and Donut, for image classification tasks on two different datasets. LayoutLM was trained using pre-trained weights from Microsoft, while Donut used pre-trained weights from Huggingface. Both models were fine-tuned for 100 epochs with early stopping technique, using the Adam optimizer and Cross Entropy Loss. Our results show that LayoutLM performs better than Donut on the first dataset, achieving an overall accuracy of 0.88, while Donut achieved an accuracy of 0.74. Our study demonstrates the importance of carefully selecting and evaluating different models for document classification tasks, based on the specific char- acteristics of the dataset and the task requirements. Additionally, we provide insights into the strengths and weaknesses of both LayoutLM and Donut models for document classification on different datasets.
Some of the metrics are blocked by your
consent settings
Wireless sensor networks localization methods: Multidimensional scaling vs. semidefinite programming approach
(Springer, Berlin, Heidelberg, 2009-09-28)
;
Davcev, Danco
;
With the recent development of technology, wireless sensor networks are becoming an important part of many applications such as health and medical applications, military applications, agriculture monitoring, home and office applications, environmental monitoring, etc. Knowing the location of a sensor is important, but GPS receivers and sophisticated sensors are too expensive and require processing power. Therefore, the localization wireless sensor network problem is a growing field of interest. The aim of this paper is to give a comparison of wireless sensor network localization methods, and therefore, multidimensional scaling and semidefinite programming are chosen for this research. Multidimensional scaling is a simple mathematical technique widely-discussed that solves the wireless sensor networks localization problem. In contrast, semidefinite programming is a relatively new field of optimization with a growing use, although being more complex. In this paper, using extensive simulations, a detailed overview of these two approaches is given, regarding different network topologies, various network parameters and performance issues. The performances of both techniques are highly satisfactory and estimation errors are minimal.

Stojkoska, Biljana

Filters

Settings

Sort By

Results per page