Purpose: Early identification of patients at risk for acute respiratory failure (ARF) could help clinicians devise preventive strategies. Analyzing biosignals with artificial intelligence (AI) can uncover hidden information and variability within time series. We aimed to develop and validate AI models to predict ARF within 72 h after emergency department admission, primarily using highresolution biosignals collected within 4 h of arrival. Materials and Methods: Our AI model, built on convolutional recurrent neural networks, combines biosignal feature extraction and sequence modeling. The model was developed and internally validated with data from 5284 admissions [1085 (20.5%) positive for ARF], and externally validated using data from 144 admissions [7 (4.9%) positive for ARF] from another institution. We defined ARF as the application of advanced respiratory support devices. Results: Our AI model performed well in predicting ARF, achieving area under the receiver operating characteristic curve (AUROC) of 0.840 and 0.743 in internal and external validations, respectively. It outperformed the Modified Early Warning Score (MEWS) and XGBoost models built only with clinical variables. High predictive ability for mortality was observed, with AUROC up to 0.809. A 10% increase in AI prediction scores was associated with 1.44-fold and 1.42-fold increases in ARF risk and mortality risk, respectively, even after adjusting for MEWS and demographic variables. Conclusion Our AI model demonstrates high predictive accuracy and significant associations with clinical outcomes. Our AI model has the potential to promptly aid in triage decisions. Our study shows that using AI to analyze biosignals advances disease detection and prediction.

Purpose: Early identification of patients at risk for acute respiratory failure (ARF) could help clinicians devise preventive strategies. Analyzing biosignals with artificial intelligence (AI) can uncover hidden information and variability within time series. We aimed to develop and validate AI models to predict ARF within 72 h after emergency department admission, primarily using highresolution biosignals collected within 4 h of arrival. Materials and Methods: Our AI model, built on convolutional recurrent neural networks, combines biosignal feature extraction and sequence modeling. The model was developed and internally validated with data from 5284 admissions [1085 (20.5%) positive for ARF], and externally validated using data from 144 admissions [7 (4.9%) positive for ARF] from another institution. We defined ARF as the application of advanced respiratory support devices. Results: Our AI model performed well in predicting ARF, achieving area under the receiver operating characteristic curve (AUROC) of 0.840 and 0.743 in internal and external validations, respectively. It outperformed the Modified Early Warning Score (MEWS) and XGBoost models built only with clinical variables. High predictive ability for mortality was observed, with AUROC up to 0.809. A 10% increase in AI prediction scores was associated with 1.44-fold and 1.42-fold increases in ARF risk and mortality risk, respectively, even after adjusting for MEWS and demographic variables. Conclusion Our AI model demonstrates high predictive accuracy and significant associations with clinical outcomes. Our AI model has the potential to promptly aid in triage decisions. Our study shows that using AI to analyze biosignals advances disease detection and prediction.

Background/Aims: The incidence of steatotic liver disease (SLD) is increasing across all age groups as the incidence of obesity increases worldwide. The existing noninvasive prediction models for SLD require laboratory tests or imaging and perform poorly in the early diagnosis of infrequently screened populations such as young adults and individuals with healthcare disparities. We developed a machine learning-based point-of-care prediction model for SLD that is readily available to the broader population with the aim of facilitating early detection and timely intervention and ultimately reducing the burden of SLD. Methods: We retrospectively analyzed the clinical data of 28,506 adults who had routine health check-ups in South Korea from January to December 2022. A total of 229,162 individuals were included in the external validation study. Data were analyzed and predictions were made using a logistic regression model with machine learning algorithms. Results: A total of 20,094 individuals were categorized into SLD and non-SLD groups on the basis of the presence of fatty liver disease. We developed three prediction models: SLD model 1, which included age and body mass index (BMI); SLD model 2, which included BMI and body fat per muscle mass; and SLD model 3, which included BMI and visceral fat per muscle mass. In the derivation cohort, the area under the receiver operating characteristic curve (AUROC) was 0.817 for model 1, 0.821 for model 2, and 0.820 for model 3. In the internal validation cohort, 86.9% of individuals were correctly classified by the SLD models. The external validation study revealed an AUROC above 0.84 for all the models. Conclusions As our three novel SLD prediction models are cost-effective, noninvasive, and accessible, they could serve as validated clinical tools for mass screening of SLD.

Purpose: Early identification of patients at risk for acute respiratory failure (ARF) could help clinicians devise preventive strategies. Analyzing biosignals with artificial intelligence (AI) can uncover hidden information and variability within time series. We aimed to develop and validate AI models to predict ARF within 72 h after emergency department admission, primarily using highresolution biosignals collected within 4 h of arrival. Materials and Methods: Our AI model, built on convolutional recurrent neural networks, combines biosignal feature extraction and sequence modeling. The model was developed and internally validated with data from 5284 admissions [1085 (20.5%) positive for ARF], and externally validated using data from 144 admissions [7 (4.9%) positive for ARF] from another institution. We defined ARF as the application of advanced respiratory support devices. Results: Our AI model performed well in predicting ARF, achieving area under the receiver operating characteristic curve (AUROC) of 0.840 and 0.743 in internal and external validations, respectively. It outperformed the Modified Early Warning Score (MEWS) and XGBoost models built only with clinical variables. High predictive ability for mortality was observed, with AUROC up to 0.809. A 10% increase in AI prediction scores was associated with 1.44-fold and 1.42-fold increases in ARF risk and mortality risk, respectively, even after adjusting for MEWS and demographic variables. Conclusion Our AI model demonstrates high predictive accuracy and significant associations with clinical outcomes. Our AI model has the potential to promptly aid in triage decisions. Our study shows that using AI to analyze biosignals advances disease detection and prediction.

Background: The accuracy of Logical Observation Identifiers Names and Codes (LOINC) mappings is reportedly low, and the LOINC codes used for research purposes in Korea have not been validated for accuracy or usability. Our study aimed to evaluate the discrepancies and similarities in interoperability using existing LOINC mappings in actual patient care settings. Methods: We collected data on local test codes and their corresponding LOINC mappings from seven university hospitals. Our analysis focused on laboratory tests that are frequently requested, excluding clinical microbiology and molecular tests. Codes from nationwide proficiency tests served as intermediary benchmarks for comparison. A research team, comprising clinical pathologists and terminology experts, utilized the LOINC manual to reach a consensus on determining the most suitable LOINC codes. Results: A total of 235 LOINC codes were designated as optimal codes for 162 frequent tests.Among these, 51 test items, including 34 urine tests, required multiple optimal LOINC codes, primarily due to unnoted properties such as whether the test was quantitative or qualitative, or differences in measurement units. We analyzed 962 LOINC codes linked to 162 tests across seven institutions, discovering that 792 (82.3%) of these codes were consistent. Inconsistencies were most common in the analyte component (38 inconsistencies, 33.3%), followed by the method (33 inconsistencies, 28.9%), and properties (13 inconsistencies, 11.4%). Conclusion This study reveals a significant inconsistency rate of over 15% in LOINC mappings utilized for research purposes in university hospitals, underlining the necessity for expert verification to enhance interoperability in real patient care.

Background: The accuracy of Logical Observation Identifiers Names and Codes (LOINC) mappings is reportedly low, and the LOINC codes used for research purposes in Korea have not been validated for accuracy or usability. Our study aimed to evaluate the discrepancies and similarities in interoperability using existing LOINC mappings in actual patient care settings. Methods: We collected data on local test codes and their corresponding LOINC mappings from seven university hospitals. Our analysis focused on laboratory tests that are frequently requested, excluding clinical microbiology and molecular tests. Codes from nationwide proficiency tests served as intermediary benchmarks for comparison. A research team, comprising clinical pathologists and terminology experts, utilized the LOINC manual to reach a consensus on determining the most suitable LOINC codes. Results: A total of 235 LOINC codes were designated as optimal codes for 162 frequent tests.Among these, 51 test items, including 34 urine tests, required multiple optimal LOINC codes, primarily due to unnoted properties such as whether the test was quantitative or qualitative, or differences in measurement units. We analyzed 962 LOINC codes linked to 162 tests across seven institutions, discovering that 792 (82.3%) of these codes were consistent. Inconsistencies were most common in the analyte component (38 inconsistencies, 33.3%), followed by the method (33 inconsistencies, 28.9%), and properties (13 inconsistencies, 11.4%). Conclusion This study reveals a significant inconsistency rate of over 15% in LOINC mappings utilized for research purposes in university hospitals, underlining the necessity for expert verification to enhance interoperability in real patient care.

Purpose: Since 1995, the Korean Gastric Cancer Association (KGCA) has been periodically conducting nationwide surveys on patients with surgically treated gastric cancer. This study details the results of the survey conducted in 2023. Materials and Methods: The survey was conducted from March to December 2024 using a standardized case report form. Data were collected on 86 items, including patient demographics, tumor characteristics, surgical procedures, and surgical outcomes. The results of the 2023 survey were compared with those of previous surveys. Results: Data from 12,751 cases were collected from 66 institutions. The mean patient age was 64.6 years, and the proportion of patients aged ≥71 years increased from 9.1% in 1995 to 31.7% in 2023. The proportion of upper-third tumors slightly decreased to 16.8% compared to 20.9% in 2019. Early gastric cancer accounted for 63.1% of cases in 2023.Regarding operative procedures, a totally laparoscopic approach was most frequently applied (63.2%) in 2023, while robotic gastrectomy steadily increased to 9.5% from 2.1% in 2014.The most common anastomotic method was the Billroth II procedure (48.8%) after distal gastrectomy and double-tract reconstruction (51.9%) after proximal gastrectomy in 2023.However, the proportion of esophago-gastrostomy with anti-reflux procedures increased to 30.9%. The rates of post-operative mortality and overall complications were 1.0% and 15.3%, respectively. Conclusions The results of the 2023 nationwide survey demonstrate the current status of gastric cancer treatment in Korea. This information will provide a basis for future gastric cancer research.

Background: The use of acellular dermal matrix (ADM) in breast reconstruction can inhibit capsular contracture, increasing the success rate of surgery. Adipose-derived stem cells (ADSCs) can effectively suppress foreign body reaction, which is a major cause of capsular contracture. This study aimed to elucidate the synergistic effects of combining ADSCs with ADM on capsule formation, utilizing a rat model. Methods: The study utilized 12 rats, equally divided into two experimental groups. Group A received silicone implants covered with ADM, while Group B was implanted with silicone prostheses wrapped in ADM, pre-seeded with ADSCs. Capsule formation was assessed through visual examination, histological analysis, and reverse transcription-polymerase chain reaction (RT-PCR) at 4 and 8 weeks post-implantation. Results: At 4 weeks, the mean capsular thickness was 177.16 μm in Group A and 170.76 μm in Group B; at 8 weeks, it was 196.69 μm in Group A and 176.10 μm in Group B. Statistical analysis showed no significant difference in capsule thickness between the groups (P>0.05). Histological findings indicated that Group A had more inflammatory cells and collagen fibers and reduced angiogenesis. RT-PCR showed that angiogenesis-promoting gene expression in Group B was 14% higher at 4 weeks and 156% higher at 8 weeks compared to Group A. Conclusion Although no statistically significant reduction in capsule thickness was observed, ADSC-seeded implants showed histological features associated with reduced inflammation and enhanced angiogenesis, suggesting potential benefits in capsule formation management.

Background/Aims: The incidence of steatotic liver disease (SLD) is increasing across all age groups as the incidence of obesity increases worldwide. The existing noninvasive prediction models for SLD require laboratory tests or imaging and perform poorly in the early diagnosis of infrequently screened populations such as young adults and individuals with healthcare disparities. We developed a machine learning-based point-of-care prediction model for SLD that is readily available to the broader population with the aim of facilitating early detection and timely intervention and ultimately reducing the burden of SLD. Methods: We retrospectively analyzed the clinical data of 28,506 adults who had routine health check-ups in South Korea from January to December 2022. A total of 229,162 individuals were included in the external validation study. Data were analyzed and predictions were made using a logistic regression model with machine learning algorithms. Results: A total of 20,094 individuals were categorized into SLD and non-SLD groups on the basis of the presence of fatty liver disease. We developed three prediction models: SLD model 1, which included age and body mass index (BMI); SLD model 2, which included BMI and body fat per muscle mass; and SLD model 3, which included BMI and visceral fat per muscle mass. In the derivation cohort, the area under the receiver operating characteristic curve (AUROC) was 0.817 for model 1, 0.821 for model 2, and 0.820 for model 3. In the internal validation cohort, 86.9% of individuals were correctly classified by the SLD models. The external validation study revealed an AUROC above 0.84 for all the models. Conclusions As our three novel SLD prediction models are cost-effective, noninvasive, and accessible, they could serve as validated clinical tools for mass screening of SLD.