https://scholar.korea.ac.kr/handle/2021.sw.korea/2830 2026-04-05T17:29:52Z 2026-04-05T17:29:52Z 홍지윤 주아림 전수영 https://scholar.korea.ac.kr/handle/2021.sw.korea/269468 2025-05-13T00:00:19Z 2025-04-01T00:00:00Z

Title: RAG와 LLM을 활용한 금융 심사 설명력 향상 방안 Authors: 홍지윤; 주아림; 전수영 Abstract: 최근 대규모 언어모델(LLM)의 발전으로 자연어 처리 기반 시스템의 활용 가능성이 크게 확장되고 있다. 특히 금융 도메인에서는 고객 맞춤형 피드백을 통한 설명 가능성과 신뢰성이 중요한 요소로 부상하고 있다. 그러나 기존의 규칙 기반 금융대출심사 시스템은 개인의 구체적 재무 상황과 맥락을 충분히 반영하지 못하고 정형화된 피드백만을 제공하여 고객의 행동 변화를 유도하는 데 한계가 있다. 본 연구는 이러한 문제점을 해결하기 위해 RAG(retrieval augmented generation)를 결합한 LLM 기반 금융대출심사 시스템을 제안한다. 제안된 시스템은 외부 금융 문서를 RAG로 검색한 뒤 검색된 컨텍스트가 LLM에 입력으로 들어가는 구조로 구현되었다. 이는 금융 문서 및 고객 데이터를 기반으로 실시간으로 외부 지식을 검색하여 전문적인 문서 기반의 구체적이고 실행 가능한 금융 개선 방안을 제공한다. 실제 고객 데이터를 활용한 실험 결과, 제안된 RAG 기반 시스템은 기존 시스템 대비 정보 검색 및 자연어 생성 성능 측면에서 의미 있는 성능 향상을 보였으며, 전문적인 금융 문서 참조를 통해 개인 맞춤형 금융 서비스의 신뢰성과 효율성을 높이는 데 기여할 수 있음을 보여준다.

2025-04-01T00:00:00Z Holm, Rochelle H. Rempala, Grzegorz A. Choi, Boseung Brick, J. Michael Amraotkar, Alok R. Keith, Rachel J. Rouchka, Eric C. Chariker, Julia H. Palmer, Kenneth E. Smith, Ted Bhatnagar, Aruni https://scholar.korea.ac.kr/handle/2021.sw.korea/199812 2024-11-19T01:05:44Z 2024-12-01T00:00:00Z

Title: Dynamic SARS-CoV-2 surveillance model combining seroprevalence and wastewater concentrations for post-vaccine disease burden estimates Authors: Holm, Rochelle H.; Rempala, Grzegorz A.; Choi, Boseung; Brick, J. Michael; Amraotkar, Alok R.; Keith, Rachel J.; Rouchka, Eric C.; Chariker, Julia H.; Palmer, Kenneth E.; Smith, Ted; Bhatnagar, Aruni Abstract: Background: Despite wide scale assessments, it remains unclear how large-scale severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) vaccination affected the wastewater concentration of the virus or the overall disease burden as measured by hospitalization rates. Methods: We used weekly SARS-CoV-2 wastewater concentration with a stratified random sampling of seroprevalence, and linked vaccination and hospitalization data, from April 2021–August 2021 in Jefferson County, Kentucky (USA). Our susceptible (S), vaccinated (V), variant-specific infected (I1 and I2), recovered (R), and seropositive (T) model (SVI2RT) tracked prevalence longitudinally. This was related to wastewater concentration. Results: Here we show the 64% county vaccination rate translate into about a 61% decrease in SARS-CoV-2 incidence. The estimated effect of SARS-CoV-2 Delta variant emergence is a 24-fold increase of infection counts, which correspond to an over 9-fold increase in wastewater concentration. Hospitalization burden and wastewater concentration have the strongest correlation (r = 0.95) at 1 week lag. Conclusions: Our study underscores the importance of continuing environmental surveillance post-vaccine and provides a proof-of-concept for environmental epidemiology monitoring of infectious disease for future pandemic preparedness. © The Author(s) 2024.

2024-12-01T00:00:00Z Hong, Hyukpyo Eom, Eunjin Lee, Hyojung Choi, Sunhwa Choi, Boseung Kim, Jae Kyoung https://scholar.korea.ac.kr/handle/2021.sw.korea/201068 2025-02-14T05:31:00Z 2024-10-09T00:00:00Z

Title: Overcoming bias in estimating epidemiological parameters with realistic history-dependent disease spread dynamics Authors: Hong, Hyukpyo; Eom, Eunjin; Lee, Hyojung; Choi, Sunhwa; Choi, Boseung; Kim, Jae Kyoung Abstract: Epidemiological parameters such as the reproduction number, latent period, and infectious period provide crucial information about the spread of infectious diseases and directly inform intervention strategies. These parameters have generally been estimated by mathematical models that involve an unrealistic assumption of history-independent dynamics for simplicity. This assumes that the chance of becoming infectious during the latent period or recovering during the infectious period remains constant, whereas in reality, these chances vary over time. Here, we find that conventional approaches with this assumption cause serious bias in epidemiological parameter estimation. To address this bias, we developed a Bayesian inference method by adopting more realistic history-dependent disease dynamics. Our method more accurately and precisely estimates the reproduction number than the conventional approaches solely from confirmed cases data, which are easy to obtain through testing. It also revealed how the infectious period distribution changed throughout the COVID-19 pandemic during 2020 in South Korea. We also provide a user-friendly package, IONISE, that automates this method. In infectious disease models, epidemiological parameters are typically estimated assuming a exponential distributions of latent and infectious periods. Here, the authors show that adapting models to incorporate Gamma distributions produces less biased estimates when applied to the example of the early COVID-19 pandemic in South Korea.

2024-10-09T00:00:00Z Kim, Hayeon Seo, Jun-Ho Nam, Jin Hyun Lim, Yejee Choi, Kyung Hee Kim, Kyungim https://scholar.korea.ac.kr/handle/2021.sw.korea/200825 2025-02-14T05:01:22Z 2024-10-01T00:00:00Z

Title: Comparing ischemic cardiovascular effectiveness and safety between individual SGLT-2 inhibitors and DPP-4 inhibitors in patients with type 2 diabetes: a nationwide population-based cohort study Authors: Kim, Hayeon; Seo, Jun-Ho; Nam, Jin Hyun; Lim, Yejee; Choi, Kyung Hee; Kim, Kyungim Abstract: Objectives: This study compared the ischemic cardiovascular events (iCVEs) effectiveness and safety of initiating empagliflozin or dapagliflozin with those of dipeptidyl peptidase-4 inhibitors (DPP-4is), as well as the comparative effects between empagliflozin and dapagliflozin. Methods: Using data from the National Health Insurance Service in Korea, patients with type 2 diabetes mellitus (T2DM) who were newly prescribed empagliflozin, dapagliflozin, or DPP-4is from 2016 to 2019 and who did not have a recent CVE history were included. A Cox proportional hazards regression model was used to estimate the adjusted hazard ratio (aHR) with 95% confidence intervals (CIs) for iCVEs and safety events. Results: Empagliflozin and dapagliflozin significantly reduced the risks of ischemic stroke (aHR 0.568, 95% CI 0.408–0.791; aHR 0.612, 95% CI 0.476–0.786, respectively) and all-cause mortality (aHR 0.590, 95% CI 0.442–0.788; aHR 0.730, 95% CI 0.603–0.884, respectively) compared with DPP-4is. Initiating dapagliflozin or empagliflozin was associated with significantly lower incidence of severe hypoglycemia, bone fracture, urinary tract infection, and acute kidney injury than that of DPP-4is. No significant differences were observed between empagliflozin and dapagliflozin in iCVEs and most safety outcomes. Conclusion: Empagliflozin and dapagliflozin showed significant preventive effects on ischemic stroke and all-cause mortality compared with DPP-4is in patients with T2DM, and their protective effects were similar. Both empagliflozin and dapagliflozin were not related to the harmful effects on most safety events. These results suggest that it may be beneficial to initiate empagliflozin or dapagliflozin for ischemic stroke prevention in patients with T2DM. However, further validation studies, such as randomized controlled trials, are needed to generalize these results. Copyright © 2024 Kim, Seo, Nam, Lim, Choi and Kim.

2024-10-01T00:00:00Z