https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/263 2026-07-04T01:53:37Z https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/213851 Title: Prediction of Retinopathy of Prematurity and Treatment in Very Low Birth Weight Infants Using Machine Learning on Nationwide Non-Imaging Clinical Data Authors: Hwang, Jae Kyoon; Jung, Donggoo; Park, Hyun-Kyung; Kim, Daehyun; Do, Hyun Jeong; Oh, Seong Hee; Kim, Seung Hyun; Kim, Tae Hyun; Jin, Hyunseung Abstract: Introduction: Retinopathy of prematurity (ROP) remains a leading cause of preventable blindness in preterm infants. This study aimed to develop machine learning (ML) models using non-imaging clinical data to predict ROP, severe ROP (sROP), and treated ROP (tROP) in very low birth weight (VLBW) infants. Methods: We utilized nationwide clinical data from the Korean Neonatal Network, including 44 perinatal and neonatal variables. Two deep learning models, Multilayer Perceptron (MLP) and Neural Oblivious Decision Ensembles (NODE), optimized for tabular data, were applied. Additionally, we developed simplified models using eight key variables selected through clinical and algorithmic relevance. Results: MLP and NODE models demonstrated high predictive performance. For the full 44-variable models, the area under the receiver operating characteristic curve (AUROC) was as follows: ROP (0.853/0.855), sROP (0.888/0.890), and tROP (0.905/0.909). The reduced 8-variable models yielded comparable AUROCs: ROP (0.851/0.855), sROP (0.895/0.895), and tROP (0.910/0.909). Conclusion: The proposed ML models based on nationwide non-imaging clinical data enable early risk identification and timely intervention for ROP in VLBW infants. This cost-effective and scalable approach may help improve outcomes, especially in resource-limited settings. Retinopathy of prematurity (ROP) is an eye condition that can affect premature babies (babies born too early, before 37 weeks of pregnancy). In ROP, abnormal blood vessels grow in the retina, which can lead to vision problems or even blindness. To prevent serious outcomes, early detection and treatment are essential. However, not all hospitals have enough trained eye specialists to screen every baby at risk. For this reason, this study aimed to develop an easier way to identify babies who may need eye examinations using commonly collected medical data. To address this goal, the researchers analyzed health records of premature babies collected across South Korea. Using a method called machine learning, which allows computers to find patterns in data, they created two computer models. These models could predict which babies were more likely to develop severe forms of ROP or need treatment. Importantly, the models used only basic clinical information like birth weight, oxygen support, and medical complications, without requiring eye images. The models showed high accuracy even when using just a few key factors. By identifying risk in this way, this type of model can help hospitals recognize high-risk babies early and refer them for specialized care, even if eye doctors are not available on site. It offers a practical, low-cost tool for improving ROP screening programs, especially in areas with limited resources. 2026-06-01T00:00:00Z https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/217789 Title: DA-BioNER: data augmentation based on few-shot learning and distant supervision for biomedical named entity recognition Authors: Park, Yesol; Son, Gyujin; Kim, Taeuk; Rho, Mina Abstract: Motivation Named entity recognition (NER) is a fundamental component of structured knowledge extraction, yet its effectiveness in emerging domains remains by the scarcity of high-quality, domain-specific annotated corpora. Although data augmentation and distant supervision have been explored to alleviate this issue, existing methods often introduce limited entity diversity, noisy labels, or disrupt contextual integrity, thereby limiting their generalization ability in low-resource settings.Results In this study, we propose DA-BioNER, a context-preserving data expansion framework for biomedical NER. DA-BioNER combines multiple base NER models trained on few-shot data to provide coarse annotations, followed by refinement using a large language model (LLM) guided by global biomedical knowledge. Unlike generation-based augmentation methods that synthesize new sentences, DA-BioNER performs annotation refinement within existing sentences, preserving both syntactic structure and semantic context. By constraining the role of LLM to refinement rather than open-ended generation, the framework effectively reduces hallucination while improving label precision and consistency. We evaluate DA-BioNER on three benchmark datasets (NCBI-Disease, BC5CDR, and BioRED), under low-resource conditions. In 40-shot settings, DA-BioNER achieves F1-scores of 0.750, 0.795, and 0.799, respectively, outperforming state-of-the-art methods, including LSMS, DAGA, and MELM, by up to 0.32. Under more extreme few-shot settings, DA-BioNER further improves F1-scores by up to 0.08, while generating an average of 1,391 additional unique entities, substantially enriching training diversity. These results demonstrate that DA-BioNER provides a scalable and adaptable solution for robust biomedical NER, particularly in domain adaptation and low-resource scenarios.Availability DA-BioNER is publicly available at https://github.com/DMnBI/DA-BioNER. 2026-06-01T00:00:00Z https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/217787 Title: Cache-aware stream consolidation in a server cluster for OTT VoD services Authors: Kim, Eunsam; Lee, Choonhwa Abstract: The global proliferation of Over-the-Top (OTT) streaming services has provided unprecedented access to extensive high-definition video content across diverse devices and geographic regions, creating immense pressure on data center infrastructures. To handle the resulting surge in video traffic, OTT providers typically employ hierarchical architectures comprising origin data centers and geographically distributed Content Delivery Network (CDN) edge data centers, where both are structured as clusters of multiple servers with limited storage and processing capacity. To ensure high availability and adequate service capacity for popular videos in a server cluster, traditional approaches rely heavily on replication and load balancing strategies. However, these strategies often lead to significant inefficiencies in resource utilization due to excessive replica creation and degraded caching performance caused by stream fragmentation. In contrast, interval caching not only significantly improves caching performance by exploiting temporal locality between streams, but also reduces the need for replication by conserving resources. To fully exploit these benefits, we focus on stream consolidation as a means to create additional caching opportunities by reducing the temporal gaps between consecutive requests. To this end, we propose a novel cache-aware stream consolidation scheme that dynamically performs aggregation, replication, de-replication, and request steering operations based on real-time system load and the degree of stream consolidation for each video across servers. The main contributions of our proposed scheme are two-fold: First, it enhances resource utilization efficiency by applying an aggregation-first policy before resorting to replication, significantly reducing unnecessary replications while still maintaining adequate service capacity. Second, it improves caching efficiency by consolidating streams for the same video onto fewer servers, substantially shortening intervals between consecutive requests. Through extensive experiments, we demonstrate that our proposed scheme significantly outperforms conventional schemes including replication-only and replication with interval caching under varying parameters. 2026-06-01T00:00:00Z https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/212941 Title: Reducing Cybersickness for 2D VR Spectators Using a Gaze-Based Stabilized Third-Person View Authors: Kim, Sungnam; Park, Jong-Il Abstract: While motion sickness in VR has been widely studied for HMD (head-mounted display) wearers, the experience of external spectators watching VR content on 2D monitors has received relatively little attention. This paper proposes a gaze-based stabilized thirdperson view that adjusts the spectator camera according to the HMD gaze behavior of the HMD wearer. Through a repeated measures experiment using the Simulator Sickness Questionnaire (SSQ), we show that the proposed view reduces cybersickness experienced by spectators watching VR content on 2D monitors, compared to conventional first-person and third-person spectator views. 2026-05-01T00:00:00Z