Wona Choi

Background: Real-world data (RWD)–based pediatric acute lymphoblastic leukemia (ALL) registries often face scalability and reliability limits owing to manual data entry and inter-institutional heterogeneity. We evaluated the feasibility of building an automated, high-quality registry using electronic medical records (EMR) and clinical data warehouses (CDWs) to quantitatively assess the Pediatric ALL Automated Registry (PeARL) for (1) automatic extractability (ADE) (2), data quality, and (3) multicenter applicability. Methods: In this study, we used CDWs to identify patients aged < 18 years with ALL (ICD-10 C91.0) who visited Seoul National University Hospital (SNUH) and the Catholic Medical Center (CMC). between 1990 and 2023. An automated extraction pipeline applied standard mapping, multivariate transformations, and rule-based natural language processing (NLP). Key variables were selected using the Medical Information Standards for Hematologic Cancer. Data quality was evaluated based on 228 rules across five dimensions (completeness, validity, accuracy, uniqueness, and consistency) based on DQ4HEALTH; error rates, defined as the proportion of data elements violating these rules, were calculated before and after quality management. Results: Overall, 1,609 patients were included (CMC 946; SNUH 663). ADE for key variables was 89.7% at SNUH and 75.0% at CMC; most automations were single-field transformations (SNUH 61.8%; CMC 84.7%), with multivariate transformations and rule-based NLP addressing complex elements. Initial overall error rates were 1.858% (SNUH) and 0.129% (CMC), decreasing to 0.001% at both institutions after quality processes. Differences in CDW structure required additional preprocessing for laboratory and transplant variables, but harmonization was achieved via standardized table specifications and cross-site review. Conclusions: PeARL integrated standardized mapping, multivariate transformations, and rule-based NLP to enable large-scale automation of a pediatric ALL registry, achieving a 0.001% overall error rate under DQ4HEALTH-based quality management. This clinically guided, standardized framework enables reproducible implementation and scalable automation of pediatric ALL registry construction, supporting multicenter research and the generation of regulatory-grade real-world evidence. Clinical trial number: Not applicable.

Background: This study addresses the gap in knowledge regarding the long-term mortality implications of postoperative acute kidney injury (PO-AKI) utilizing advanced machine learning techniques to predict outcomes more accurately than traditional statistical models.Methods: A retrospective cohort study was conducted using data from seven institutions between March 2009 and December 2019. Machine learning models were developed to predict all-cause mortality of PO-AKI patients using 23 preoperative variables and one postoperative variable. Model performance was compared to a traditional statistical approach with Cox regression analysis. The concordance index was used as a predictive performance metric to compare prediction capabilities among different models.Results: Among 199,403 patients, 2,105 developed PO-AKI. During a median follow-up of 144 months (interquartile range, 99.61–170.71 months), 472 in-hospital deaths occurred. Subjects with PO-AKI had a significantly lower survival rate than those without PO-AKI (p < 0.001). For predicting mortality, the XGBoost with an accelerated failure time model had the highest concordance index (0.7521), followed by random survival forest (0.7371), multivariable Cox regression model (0.7318), survival support vector machine (0.7304), and gradient boosting (0.7277). Conclusion: XGBoost with an accelerated failure time model was developed in this study to predict long-term mortality associated with PO-AKI. Its performance was superior to conventional models. The application of machine learning techniques may offer a promising approach to predict mortality following PO-AKI more accurately, providing a basis for developing targeted interventions and clinical guidelines to improve patient outcomes.

Background: Advanced biobanks increasingly focus on supporting biomedical research through the collection and integration of large-scale biological and clinical datasets. This study aimed to develop a big data platform that enables institutions within the Korea Biobank Network (KBN) to efficiently collect and utilize clinical information using a standardized common data model. Methods: The KBN Biobank Research Information and Digital Image Exchange (BRIDGE) platform was developed to allow 43 biobanks to systemically collect and upload electronic medical records and clinical data. This platform was designed to incorporate automated quality verification and basic statistical preprocessing functionalities, allowing users to analyze data efficiently without complex queries. Additionally, a survey was conducted to evaluate user satisfaction with the platform. Results: Through the KBN BRIDGE platform, institutions collected and integrated clinical information on 39 diseases. A total of 136,473 patients’ clinical data, collected by institutions between 2021 and 2023, were uploaded to the KBN common data model, including 43,330 serum samples, 33,352 plasma samples, and 22,279 buffy coat samples. A satisfaction survey conducted among 35 institutional data managers reported an average score of 3.5 out of 5 for the platform. Conclusions: This study developed and demonstrated that the KBN BRIDGE platform enables institutions to systematically collect, integrate, and manage large-scale clinical information across multiple biobanks. Furthermore, through data quality management and preprocessing statistical functions, the platform has shown potential for several research applications. Future improvements in system functionality and clinical information utilization can further enhance the platform’s utility across various research fields.

This study aims to develop a Clinical Decision Support System (CDSS) platform utilizing healthcare MyData to enhance patient safety. The platform focuses on algorithms for managing hyperglycemia and predicting Acute Kidney Injury (AKI). To evaluate the platform, we conducted in-depth interviews with healthcare professionals and assessed its performance. By integrating artificial intelligence (AI) with MyData data, this research seeks to deliver personalized safety alerts, marking progress in patient-centered care through the application of healthcare MyData.

This study aims to develop predictive models for diabetes and hypertension onset using personal health records and deep learning to analyze time-series data patterns. By leveraging the MyData initiative, the model enhances early detection and prevention strategies through personalized prediction. Results focus on evaluating performance using AUROC, F1-score, and Recall, highlighting its potential for accurate chronic disease prediction. Despite requiring further generalization, this framework lays the foundation for real-world applications to improve preventive care and advance personalized healthcare.

This study evaluates the performance of the Llama 3.1 model with 70 billion parameters (70b) in extracting clinically significant information from free-text pathology reports under zero-shot learning conditions. The model achieved a remarkable accuracy of 98%, directly generating structured outputs in JSON format without requiring annotated examples. These findings demonstrate the potential of Llama 3.1 in automating data extraction from complex medical documents, streamlining clinical workflows, and enhancing the precision of information retrieval in healthcare settings.

The KBN (Korea Biobank Network) BRIDGE (Biobank Research Information and Digital image Exchange) platform enables Korean biobanks to standardize and upload clinical data based on the KBN Common Data Model (CDM), and allows access to data from their own institution or consortia led by their institution. This study assessed user acceptance using the Technology Acceptance Model and Information Systems Success Model. Information quality and perceived usefulness showed strong positive correlations with satisfaction. Experienced users reported significantly higher satisfaction and service quality compared to less experienced users.

As the reliance on clinical epidemiological information from human specimens grows, so does the need for effective clinical information management systems, particularly for biobanks. Our study focuses on enhancing the Korea Biobank Network’s (KBN) system with data quality verification features. By comparing the quality of data collected before and after these enhancements, we observed a notable improvement in data accuracy, with the error rate decreasing from 0.1198% to 0.0492%. This advancement underscores the importance of robust data quality management in supporting high-quality clinical research and sets a precedent for the development of clinical information management systems.

Biobanks serve as vital repositories for human biospecimens and clinical data, promoting biomedical and clinical research. The integration of electronic health records particularly enhances research opportunities in the era of genomics and personalized medicine, improving understanding of tumor development and disease progression. Based on the Korea Biobank Network Common Data Model, it is possible to expand data collection across various diseases. We have developed an innovative big data platform designed to efficiently collect large-scale clinical information within the KBN. By implementing the system structure, data quality management processes, and basic statistical preprocessing functionalities, we have collected data from 136,473 individuals from 2021 to 2023, demonstrating the platform’s continuous and efficient data collection capabilities. Integration with hospital systems and robust quality management ensure the acquisition of high-quality data.

This study evaluates the efficacy of a small large language model (sLLM) in extracting critical information from free-text pathology reports across multiple centers, addressing the challenges posed by the narrative and complex nature of these documents. Employing three variants of the Llama 2 model, with 7 billion, 13 billion, and 70 billion parameters, the research assesses model performance in both zero-shot and five-shot settings, offering insights into the impact of example-based learning. A specialized information extraction tool utilizing regular expressions for pattern identification serves as the benchmark for evaluating the models’ accuracy. Conducted within a hospital’s internal environment, the study emphasizes the clinical applicability of these findings. The results reveal significant variations in model performance, with the 70 billion parameter model achieving remarkable accuracy in the five-shot scenario, demonstrating the potential of sLLMs in enhancing the efficiency and accuracy of data extraction from pathology reports. The study highlights the importance of example-driven learning and the trade-offs between model size, accuracy, hallucination rates, and processing time. These findings contribute to the ongoing efforts to integrate advanced language models into clinical settings, potentially transforming patient care and biomedical research by mitigating the limitations of manual data extraction processes.