Serpil Ustebay

Serpil Ustebay

Auto-generated profile, based on publicly available information

Serpil Ustebay

İstanbul Üniversitesi-Cerrahpaşa

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
18 September 2022 in Internal and Emergency Medicine
Verified
Journal Article
Open access
PDF file
SUSerpil Ustebay
Serpil Ustebay
ASAbdurrahman Sarmis
Abdurrahman Sarmis
GKGulsum Kubra Kaya
Gulsum Kubra Kaya
MSMark Sujan
Mark Sujan
ML algorithms are used to develop prognostic and diagnostic models and so to support clinical decision-making. This study uses eight supervised ML algorithms to predict the need for intensive care, intubation, and mortality risk for COVID-19 patients. The study uses two datasets: (1) patient demographics and clinical data (n = 11,712), and (2) patient demographics, clinical data, and blood test results (n = 602) for developing the prediction models, understanding the most significant features, and comparing the performances of eight different ML algorithms. Experimental findings showed that all prognostic prediction models reported an AUROC value of over 0.92, in which extra tree and CatBoost classifiers were often outperformed (AUROC over 0.94). The findings revealed that the features of C-reactive protein, the ratio of lymphocytes, lactic acid, and serum calcium have a substantial impact on COVID-19 prognostic predictions. This study provides evidence of the value of tree-based supervised ML algorithms for predicting prognosis in health care.

24 January 2024 in Pattern Recognition Letters
Verified
Journal Article
Closed access
BHBetul Hicdurmaz
Betul Hicdurmaz
NCNurullah Calik
Nurullah Calik
SUSerpil Ustebay
Serpil Ustebay

16 October 2024
Verified
Conference Paper
Closed access
SUSerpil Ustebay
Serpil Ustebay
BABerhan Berk Akgün
Berhan Berk Akgün
PGPiotr Gaj
Piotr Gaj
This paper presents an approach to enhancing the security of Supervisory Control and Data Acquisition (SCADA) systems through the development of Explainable Intrusion Detection Systems (X-IDS). SCADA systems play a crucial role in managing industrial processes and critical infrastructure, yet they are increasingly targeted by cyberattacks, posing significant risks to operational integrity. Leveraging Machine Learning (ML) techniques, particularly focused on network protocols involved in communication with SCADA HMI, this study proposes a protocol-based layered IDS framework to mitigate potential threats. Evaluation of the proposed models demonstrates promising results, achieving high accuracy rates surpassing previous studies. The performance metrics, including accuracy, precision, recall, and F1 scores, along with Brier scores, are comprehensively analyzed to gauge the effectiveness of the models. Additionally, Explainable Machine Learning (XAI) techniques such as SHAP values provide transparent insights into the model's decision-making process. The study highlights the importance of securing HMIs within SCADA systems and offers valuable insights for future research directions in enhancing overall network security.

26 June 2025 in Electrica
Verified
Journal Article
Open access
PDF file
SÜSerpil Üstebay
Serpil Üstebay
The Internet of Things (IoT) devices perform critical functions such as sensitive data collection, storage, and processing, which make them vulnerable to malicious attacks. In this study, a Network Intrusion Detection System was designed to enhance the security of IoT devices. Data sets obtained from three different IoT environments (CICEVSE2024, CICIoT2023, and RT-IoT2022) were utilized for attack detection using tree-based machine learning methods. Experimental results demonstrated that attacks were detected with an average accuracy of 99%. Additionally, a second security layer was implemented to identify zero-day attacks. Analyses showed that the Isolation Forest algorithm detected zero-day attacks with accuracies ranging from 30% to 62%. This proposed approach shows promise in enhancing security against known and unknown attacks. Cite this article as: S. Üstebay, “Improving zero-day attack detection accuracy in IoT networks with isolation forest and tree-based models,” Electrica, 25, 0177, 2025. doi: 10.5152/electrica.2025.24177.

1 February 2026 in Risk Management and Healthcare Policy
Verified
Journal Article
Open access
ASAbdurrahman Sarmis
Abdurrahman Sarmis
SUSerpil Ustebay
Serpil Ustebay
MMMuhammed Mutlu
Muhammed Mutlu
FCfurkan adem Canbaz
furkan adem Canbaz
GKGulsum Kubra Kaya
Gulsum Kubra Kaya
Urinary tract infections (UTIs) are among the most common bacterial infections worldwide, with Escherichia coli and Klebsiella pneumoniae being the predominant pathogens. Diagnostic delays necessitate the use of broad-spectrum antibiotics, which fuels antimicrobial resistance. This study aimed to develop and validate an artificial intelligence (AI) model for the rapid identification of E. coli and K. pneumoniae colonies from urine culture images. We analyzed 1547 chromogenic agar urine culture images (850 E. coli, 697 K. pneumoniae) obtained using an automated Copan WASP system. A YOLOv12 deep learning model was trained on expert-labeled colonies. The reference standard comprised MALDI-TOF Mass Spectrometry (MS) for K. pneumoniae and a characteristic chromogenic morphology for E. coli. This phenotypic method is standard in clinical practice, with validation studies for the specific agar used reporting a sensitivity of 98.1–99.0% and a specificity of 99.1% for E. coli identification. Performance was assessed on an internal test set and via external validation using 91 independent images. The model achieved 99% accuracy (precision, recall, F1-score: 0.99) on internal testing. External validation yielded 100% accuracy, with the critical note that species labels in this independent set were inferred from colony colour. Rare errors involved atypical (eg, gold-pigmented) E. coli colonies. YOLOv12 outperformed five benchmark deep learning models. This AI model enables rapid (sub-second), accurate phenotypic classification of the most common UTI pathogens directly from routine culture plates. Integration into automated systems could reduce the diagnostic timeline by approximately 18 hours, facilitating earlier targeted therapy and supporting antimicrobial stewardship. A key consideration for implementation is the model’s basis in phenotypic identification, which aligns with standard laboratory workflow but requires awareness of rare chromogenic variants. The reliance on chromogenic morphology for E. coli ground truth has been a key limitation of the study.

19 April 2026 in Düzce Üniversitesi Bilim ve Teknoloji Dergisi
Verified
Journal Article
Closed access
SÜSerpil Üstebay
Serpil Üstebay
Determining the location of moving objects in indoor environments has become a crucial requirement for many application domains. While this could be vital in hospitals, it can also enhance cultural interaction in environments such as museums. It serves a variety of usage scenarios for shopping malls, museums, and airports. Fingerprinting-based location estimation, thanks to the widespread availability of Wi-Fi access points found in nearly every building, has become one of the most common solutions for indoor positioning. This study examines the performance of Wi-Fi RSSI–based regression methods for indoor localization. Using the TUJI1 multi-device dataset, six different regression models were evaluated under Euclidean distance–based performance metrics. Among these models, XGBoost regression determined the locations of moving objects with an average positioning error of 2.07 m on the test dataset and 2.04 m during training, outperforming other linear and nonlinear regression approaches. In addition, we investigated how noisy signal measurements originating from the hardware structures of different mobile devices in the test environment affect localization systems. To analyze device heterogeneity, multiple experimental scenarios were designed, including device-specific and unified models. The findings show that tree-based ensemble models provide robust and competitive performance without requiring complex deep learning architectures.