Objectives: Large language models (LLMs) demonstrate remarkable potential in healthcare communication. However, whether they can process complex, high-volume medical information, such as stroke-related content, remains insufficiently validated. This study aimed to evaluate the natural language processing capabilities of LLMs in handling such content and to develop an evaluation instrument. Methods: A survey compared educational materials generated by two LLMs (ChatGPT 4.0 and Claude 3) with neurologist-authored content on stroke. The materials were based on two clinical scenarios representing distinct stroke etiologies: cardioembolism and large-artery atherosclerosis. They were evaluated in terms of accuracy, legality, ethics, comprehensiveness, and information delivery. Scores for comprehensiveness and information delivery were compared according to participants’ agreement with the use of LLMs in healthcare. Results: ChatGPT received the highest scores across all domains, except for legality in Scenario 2. In Scenario 1, the ranking for accuracy and summarization of clinical information was, from highest to lowest, ChatGPT, Claude, and the neurologist (η2 = 0.140, p < 0.001; η2 = 0.175, p < 0.001). The same hierarchy was observed in Scenario 2 for accuracy (η2 = 0.077, p < 0.001) and summarization (η2 = 0.194, p < 0.001). Participants who agreed with the use of LLMs in healthcare assigned higher scores for the comprehensiveness (Scenario 1, p = 0.005; Scenario 2, p = 0.007) and information delivery (Scenario 1, p = 0.003; Scenario 2, p = 0.026) of ChatGPT-generated materials than participants who did not agree. Conclusions LLMs demonstrated adequate capability to convey complex content, such as stroke-related information, in an accessible and understandable manner for non-experts.

Objectives: Migrating legacy on-premise electronic health record (EHR) systems in tertiary hospitals to modern cloud-native platforms presents technical and strategic challenges. We aimed to establish an optimized roadmap for transitioning legacy monolithic systems to a microservice architecture-based cloud-native EHR (MCEHR). Methods: We conducted a 3-month strategic assessment and case study based on the modernization requirements of a global healthcare provider. The methodology incorporated semi-structured interviews with key stakeholders, including clinical informatics officers and system architects, to identify critical pain points such as .NET 4.0 end-of-support risks and performance bottlenecks during peak clinical hours. Phased hybrid migration was adopted, analyzing over 20 TB of Oracle-based legacy data and evaluating the technical feasibility of transitioning to .NET 9 and RESTful APIs. To ensure clinical safety, a proof-of-concept (PoC) environment was developed to simulate high-concurrency clinical workloads, emphasizing system resilience and transaction integrity during intensive order-entry periods. Results: The transition to .NET 9 and MCEHR demonstrated 100% transaction integrity across 1,012 complex clinical test cases. Frontend and backend modernization showed high feasibility; however, migration of business logic embedded within legacy Oracle views represented a primary technical bottleneck, necessitating targeted decoupling. The PoC confirmed that RESTful API-based services maintained stable throughput under heavy concurrent loads, significantly reducing the risk of system-induced delays in clinical workflows. Conclusions Transitioning to an MCEHR architecture is complex but strategically essential. The proposed task force team roadmap outlines staged upgrades incorporating core technology modernization (.NET 9 and RESTful APIs), selective business component migration, and parallel DevOps adoption.

Objectives: Large language models (LLMs) are increasingly being used for medical question-answering (QA) tasks. However, most models are trained primarily on English-language data, which limits their effectiveness in non-English clinical contexts. In healthcare settings such as Korea, LLMs adapted to both the local language and the medical domain are needed. This study aimed to evaluate the performance of Korean-native and multilingual LLMs fine-tuned on Korean arthritis-related medical QA data and to examine the impact of language and domain adaptation. Methods: A dataset of 5,451 Korean QA pairs related to arthritis was constructed from a public medical corpus (AI Hub). Five LLMs (Mi:dm, EXAONE, Kanana, HyperCLOVAX, and LLaMA) were fine-tuned under identical conditions using Low-Rank Adaptation with 4-bit quantization. Model performance was evaluated on 597 validation samples using BERTScore-F1 and SBERT similarity, along with a qualitative evaluation of clinical correctness, safety, and response completeness. Results: EXAONE and HyperCLOVAX showed comparable quantitative performance in semantic accuracy and contextual consistency. Mi:dm achieved lower similarity-based scores than EXAONE and HyperCLOVAX but showed the highest performance in the qualitative evaluation, particularly for clinical correctness. Kanana exhibited moderate performance with limited domain adaptability. LLaMA showed the lowest performance in Korean medical QA, although it achieved the largest relative improvement, indicating challenges in adaptation to Korean clinical contexts. Conclusions Domain-specific fine-tuning and Korean-oriented model design improved performance in Korean arthritis medical QA. EXAONE and HyperCLOVAX achieved the highest semantic similarity, whereas Mi:dm demonstrated superior clinical correctness and safety in the qualitative evaluation. General multilingual LLaMA remained limited despite substantial gains, supporting the development of disease-specific Korean medical LLMs.

Objectives: Standardizing medication concepts across heterogeneous vocabularies is essential for interoperable analytics and observational research. In the Observational Medical Outcomes Partnership (OMOP) Common Data Model, local drug codes must be mapped to standardized RxNorm concepts, but automated mapping is challenging because drug strings encode clinically critical attributes, including strength, dosage form/route, release characteristics, and brand. Methods: We propose THIRAWAT (Terminology Harmonization using Late-Interaction Reranker With Alignment-tuned Transformers), a fine-tuned ColBERTv1 late-interaction reranker, and embed it within THIRAWAT Mapper, a retrieval–reranking pipeline with deterministic tie-breaking and stable ordering. Candidate generation used approximate nearest-neighbor retrieval with a bi-encoder (SapBERT-XLMR or BioLORD-2023). Candidates were reranked by THIRAWAT models that were fine-tuned using one-sided MaxSim and scored at inference using our adapted Bidirectional MaxSim (BiMaxSim) pooling. Finally, a deterministic tie-breaker extracted clinically salient cues, including strength, dosage form/route, release characteristics, and bracketed brand annotations, to resolve near-ties reproducibly. Results: We evaluated three mapping settings: Branded Drugs, Clinical Drugs, and Thai Medicines Terminology (TMT). Using SapBERT-XLMR retrieval with THIRAWAT-Sap-BERT reranking and deterministic tie-breaking, THIRAWAT Mapper achieved MRR@100 values of 0.954 (95% confidence interval [CI], 0.921–0.983), 0.898 (95% CI, 0.866–0.925), and 0.912 (95% CI, 0.891–0.931), outperforming a lexical term frequency–inverse document frequency baseline (0.491, 0.216, and 0.143, respectively). Hits@1 improved to 0.942 (95% CI, 0.899–0.978), 0.859 (95% CI, 0.817–0.898), and 0.868 (95% CI, 0.838–0.896), respectively. Conclusions BiMaxSim and deterministic tie-breaking improved drug mapping to RxNorm while preserving an efficient runtime profile and stable ordering. Overall, THIRAWAT Mapper offers a pragmatic combination of learned semantic matching and deterministic lexical constraints. Models and code are available on Hugging Face (https://huggingface.co/collections/sidataplus/thirawat) and GitHub (https://github.com/sidataplus/THIRAWAT-mapper).

Objectives: This study evaluated the feasibility and performance of a multi-agent (MA) system designed to support early sepsis management in intensive care units. The system integrates three specialized agents—sepsis management, antibiotic recommendation, and guideline compliance—to provide evidence-based recommendations at T = 0 hours (before culture results), extending prior single-case findings across 10 diverse cases. Methods: The MA system was powered by Palmyra- Med 70B (selected for superior MedQA performance [average score, 85.9]) and compared with GPT-3.5 Turbo and GPT-4o mini (all at a temperature of 0.25). It used retrieval-augmented generation (RAG) with ChromaDB (2021 Surviving Sepsis Campaign, over 20 high-impact manuscripts [reviews published 2018–2025] on sepsis etiologies, and other relevant sources). Eight cases from the MIMIC-IV demo and two cases from the literature were formatted as vignettes. RAG used the BAAI/bge-base-en-v1.5 embedding model with cosine similarity (threshold, 0.75) and top-5 chunks. Performance was assessed via TruLens (groundedness, approximately 0.62) and by two intensivists using a standardized questionnaire. Results: The system generated guideline-compliant recommendations (e.g., prompt surgical debridement plus meropenem and vancomycin for necrotizing fasciitis). Hallucinations occurred in three of 10 cases (e.g., “altered mental status”). Expert agreement was quantified by a Cohen kappa of 0.26. Programmatic and expert assessments showed negligible correlation. Conclusions In this exploratory study, the MA system shows preliminary promise for early sepsis support but requires human oversight to mitigate hallucinations. Code is available in GitHub; further validation is needed.

Objectives: This study aimed to evaluate and compare the accuracy of multiple artificial intelligence (AI) models (ChatGPT 5.2 Pro, Gemini 3 Fast, Claude 4.5 Sonnet, and Microsoft Copilot) in detecting orthodontic malocclusion features in standardized multiview intraoral photographs. The reference standard was assessment by an orthodontist. Methods: A cross-sectional observational study was conducted using five standardized intraoral photographs (frontal, right lateral, left lateral, maxillary occlusal, and mandibular occlusal) obtained from 50 children aged 9–12 years. The following eight malocclusion parameters were assessed: anterior crowding, diastema, overjet, overbite, molar relationship, canine relationship, crossbite, and dental arch symmetry. Diagnostic accuracy and agreement between each AI model and the orthodontist were evaluated using Cohen’s kappa (κ) and the area under the receiver operating characteristic curve (AUC). Results: Agreement between the AI models and the orthodontist ranged from poor to moderate across all orthodontic domains, with Cohen’s κ values ranging from -0.15 to 0.63. Visually prominent alignment features, including anterior crowding and diastema, demonstrated comparatively higher agreement (κ, 0.00–0.63) and discriminatory performance, with AUC values ranging from 0.56 to 0.85. In contrast, parameters requiring precise spatial interpretation, such as sagittal relationships, overbite, crossbite, and arch morphology, showed consistently low agreement (κ, -0.15 to 0.38) and poor to near-random classification performance, with AUC values predominantly ranging from 0.41 to 0.70 and, in some cases, approaching 0.50. Conclusions Current multimodal AI models demonstrate limited, parameter-dependent accuracy in detecting orthodontic malocclusions from intraoral photographs. These findings emphasize the limitations of general-purpose AI systems for orthodontic decision support and highlight the need for task-specific models trained on clinically annotated datasets.

Objectives: This study aimed to identify predictors of health literacy (HL), focusing on nonlinear relationships and interaction effects in a nationally representative population. Methods: This cross-sectional study analyzed data from 8,630 Korean adults participating in the Korea Health Panel Survey. HL was assessed using the Korean version of the European Health Literacy Survey Questionnaire 16-item (HLS-EU-Q16) and categorized as sufficient or insufficient. An extreme gradient boosting algorithm (XGBoost) was applied, incorporating survey weights. Model performance was evaluated using standard metrics, including the area under the receiver operating characteristic curve (AUC) and Brier score. Shapley additive explanations (SHAP) values were calculated to quantify individual feature importance and identify interaction effects among 69 features. Results: The XGBoost model achieved good discrimination (AUC = 0.840) and calibration (Brier score = 0.161). Age (24.4%) and education level (19.0%) were the most influential predictors. SHAP interaction analysis identified a meaningful interaction between age and education level (mean |interaction value| = 0.078), with interaction plots indicating positive patterns among adults aged 60–80 with lower educational attainment. The contribution of online health information use varied by age, showing a negative association among younger adults but a positive association among older adults. Conclusions HL is shaped by nonlinear and interactive effects across sociodemographic and health-related factors. Explainable machine learning approaches can facilitate the identification of high-priority populations and support the development of tailored, patient-centered educational strategies to improve HL and promote engagement with care.

Objectives: This study evaluated the feasibility of voice recognition-based electronic medical record (EMR) documentation for periodontal probing in dentistry, particularly emphasizing Korean-English bilingual speech patterns and real-world clinical conditions. Methods: Experiments were conducted in a dental chair setting during routine clinical hours. Environmental noise levels were measured, and two microphone types (stationary and pin-type) were evaluated. Periodontal probing phrases composed of three-digit numbers and positional terms were used for speech recognition. Consistent with common clinical practice in Korea, numerical values were spoken in Korean, whereas positional terms were spoken in English. Two speech-to-text application programming interfaces, Google Cloud Speech-to-Text and Naver Clova Speech Recognition, were assessed. Recognition accuracy was evaluated for both numerical components and complete bilingual phrases. Results: The mean environmental noise level was 60.65 dB and was minimally influenced by activity at adjacent dental chairs. The stationary microphone failed to capture speech effectively, whereas the pin-type microphone demonstrated stable recognition performance. For three-digit number recognition, accuracy was 88.3% with Google and 96.8% with Naver. For full-phrase recognition, complete matching was achieved in 36.7% of cases for Google and 52.5% for Naver. Partial recognition occurred more frequently for numerical components than for English positional terms. Conclusions Voice recognition-based EMR documentation for periodontal probing demonstrated preliminary feasibility in a dental clinical environment; however, performance was influenced by Korean-English bilingual speech patterns. These findings suggest that bilingual speech characteristics should be considered when implementing voice recognition systems in dental EMR workflows. Further optimization is required before routine clinical application.