Junmo Kim

All Publications

• MedRep: medical concept representations for general electronic health record foundation models. Journal of the American Medical Informatics Association : JAMIA Kim, J., Lee, N., Kim, J., Kim, K. 2026

  Abstract

  Traditional electronic health record (EHR) foundation models fail to process unseen medical codes, limiting generalizability across institutions with different vocabularies. To address this problem, we introduce medical concept representation (MedRep), standardized medical concept representations for EHR foundation models, enabling recognition of semantically similar concepts regardless of their specific IDs.We utilized Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) vocabulary covering 7.5 million concepts from 66 medical vocabularies. MedRep integrates large language model-generated concept descriptions and OMOP graph ontology using graph contrastive learning with knowledge distillation. We evaluated MedRep-based models on MIMIC-IV (internal validation) and EHRSHOT (external validation) across 9 prediction tasks including clinical outcomes, phenotypes, and in-hospital events.MedRep consistently outperformed baseline models, particularly in external validation with average improvements of 0.088 in area under the receiver operating characteristic curve and 0.208 in area under the precision-recall curve. Qualitative analysis demonstrated that MedRep-based models identified more clinically relevant concepts when making decisions than the baseline models. Performance improvements remained stable across diverse EHR foundation model architectures, including BEHRT, Med-BERT, and CDM-BERT.MedRep improves the generalizability of EHR foundation models by encouraging similar concepts to have similar representations. EHR foundation models developed at different institutions could cooperate through MedRep, merging knowledge from multiple hospital datasets. In addition, our approach could reduce healthcare disparities by enabling smaller institutions to benefit from models trained on larger datasets.MedRep improves EHR foundation model performance, interpretability, and generalizability, serving as a standard baseline representation for EHR foundation models adopting OMOP CDM.

  View details for DOI 10.1093/jamia/ocag032

  View details for PubMedID 41806382

• Prediction of antibiotic-associated cutaneous adverse drug reactions using electronic health record foundation models. NPJ digital medicine Kim, J., Kim, K., Yun, J. E., Hwang, Y. K., Kang, M. G., Kim, S., Yoo, S., Shin, C., Kim, S., Kim, K., Kim, S. H. 2026

  Abstract

  Cutaneous adverse drug reactions (CADRs) are the most common form of adverse drug reactions, ranging from mild rashes to life-threatening diseases, such as Stevens-Johnson syndrome and toxic epidermal necrolysis. However, there is no effective tool to predict antibiotic-associated CADRs. In this study, we propose an antibiotic-associated CADR prediction model using electronic health record (EHR) foundation models (FMs). EHR FMs are based on the pretraining-finetuning paradigms of language models, corresponding medical codes and their sequences to words and sentences. We included 802,131 inpatients across three tertiary hospitals in Korea, combining EHR data with nursing statements and reports to extract skin rash records. Our approach achieved the best predictive performance compared to all the other baseline models across all datasets. To enhance clinical relevance, we classified CADRs into immediate and delayed types and conducted a detailed sub-analysis. Finally, we found that properly configured EHR FMs can effectively predict the risk of developing antibiotics-associated CADRs, particularly for delayed-type reactions where predictive testing options are limited.

  View details for DOI 10.1038/s41746-026-02503-x

  View details for PubMedID 41775818

• Pretrained patient trajectories for adverse drug event prediction using common data model-based electronic health records. Communications medicine Kim, J., Kim, J. S., Lee, J. H., Kim, M. G., Kim, T., Cho, C., Park, R. W., Kim, K. 2025; 5 (1): 232

  Abstract

  Pretraining electronic health record (EHR) data using language models has enhanced performance across various medical tasks. Despite the potential of EHR pretraining models, predicting adverse drug events (ADEs) using EHR pretraining models has not been explored.We used observational medical outcomes partnership common data model (CDM)-based EHR data from Seoul National University Hospital (SNUH) between January 2001 and December 2023 and Ajou University Medical Center (AUMC) between January 2004 and December 2023. In total 510,879 and 419,505 adult inpatients from SNUH and AUMC are included in internal and external datasets. For pretraining, the model was trained to infer randomly masked tokens using preceding and following history. In this process, we introduced domain embedding (DE) to provide information about the domain of masked tokens, preventing the model from finding codes from irrelevant domains. For qualitative analysis, we identified important features using the attention matrix from each finetuned model.Here we show that EHR pretraining models with DE outperform the models without pretraining and DE in predicting various ADEs, with the average area under the receiver operating characteristic curve (AUROC) of 0.958 and 0.964 in internal and external validations, respectively. For feature importance analysis, we demonstrate that the results are consistent with priorly reported background clinical knowledge. In addition to cohort-level interpretation, patient-level interpretation is also available.The CDM-based EHR pretraining model with DE can improve prediction performance for various ADEs and can provide proper explanation at cohort and patient level. Our model has the potential to serve as a foundation model due to its strong prediction performance, interpretability, and compatibility.

  View details for DOI 10.1038/s43856-025-00914-7

  View details for PubMedID 40514403

  View details for PubMedCentralID PMC12166071

• Deep learning-based prediction of Clostridioides difficile infection caused by antibiotics using longitudinal electronic health records. NPJ digital medicine Kim, J., Kim, J. S., Kim, S. H., Yoo, S., Lee, J. K., Kim, K. 2024; 7 (1): 224

  Abstract

  Clostridioides difficile infection (CDI) is a major cause of antibiotic-associated diarrhea and colitis. It is recognized as one of the most significant hospital-acquired infections. Although CDI can develop severe complications and spores of Clostridioides difficile can be transmitted by the fecal-oral route, CDI is occasionally overlooked in clinical settings. Thus, it is necessary to monitor high CDI risk groups, particularly those undergoing antibiotic treatment, to prevent complications and spread. We developed and validated a deep learning-based model to predict the occurrence of CDI within 28 days after starting antibiotic treatment using longitudinal electronic health records. For each patient, timelines of vital signs and laboratory tests with a 35-day monitoring period and a patient information vector consisting of age, sex, comorbidities, and medications were constructed. Our model achieved the prediction performance with an area under the receiver operating characteristic curve of 0.952 (95% CI: 0.932-0.973) in internal validation and 0.972 (95% CI: 0.968-0.975) in external validation. Platelet count and body temperature emerged as the most important features. The risk score, the output value of the model, exhibited a consistent increase in the CDI group, while the risk score in the non-CDI group either maintained its initial value or decreased. Using our CDI prediction model, high-risk patients requiring symptom monitoring can be identified. This could help reduce the underdiagnosis of CDI, thereby decreasing transmission and preventing complications.

  View details for DOI 10.1038/s41746-024-01215-4

  View details for PubMedID 39181992

  View details for PubMedCentralID PMC11344761

• Deep learning-based long-term risk evaluation of incident type 2 diabetes using electrocardiogram in a non-diabetic population: a retrospective, multicentre study. EClinicalMedicine Kim, J., Yang, H. L., Kim, S. H., Kim, S., Lee, J., Ryu, J., Kim, K., Kim, Z., Ahn, G., Kwon, D., Yoon, H. J. 2024; 68: 102445

  Abstract

  Diabetes is a major public health concern. We aimed to evaluate the long-term risk of incident type 2 diabetes in a non-diabetic population using a deep learning model (DLM) detecting prevalent type 2 diabetes using electrocardiogram (ECG).In this retrospective study, participants who underwent health checkups at two tertiary hospitals in Seoul, South Korea, between Jan 1, 2001 and Dec 31, 2022 were included. Type 2 diabetes was defined as glucose ≥126 mg/dL or glycated haemoglobin (HbA1c) ≥ 6.5%. For survival analysis on incident type 2 diabetes, we introduced an additional variable, diabetic ECG, which is determined by the DLM trained on ECG and corresponding prevalent diabetes. It was assumed that non-diabetic individuals with diabetic ECG had a higher risk of incident type 2 diabetes than those with non-diabetic ECG. The one-dimensional ResNet-based model was adopted for the DLM, and the Guided Grad-CAM was used to localise important regions of ECG. We divided the non-diabetic group into the diabetic ECG group (false positive) and the non-diabetic ECG (true negative) group according to the DLM decision, and performed a Cox proportional hazard model, considering the occurrence of type 2 diabetes more than six months after the visit.190,581 individuals were included in the study with a median follow-up period of 11.84 years. The areas under the receiver operating characteristic curve for prevalent type 2 diabetes detection were 0.816 (0.807-0.825) and 0.762 (0.754-0.770) for the internal and external validations, respectively. The model primarily focused on the QRS duration and, occasionally, P or T waves. The diabetic ECG group exhibited an increased risk of incident type 2 diabetes compared with the non-diabetic ECG group, with hazard ratios of 2.15 (1.82-2.53) and 1.92 (1.74-2.11) for internal and external validation, respectively.In the non-diabetic group, those whose ECG was classified as diabetes by the DLM were at a higher risk of incident type 2 diabetes than those whose ECG was not. Additional clinical research on the relationship between the phenotype of ECG and diabetes to support the results and further investigation with tracked data and various ECG recording systems are suggested for future works.National Research Foundation of Korea.

  View details for DOI 10.1016/j.eclinm.2024.102445

  View details for PubMedID 38333540

  View details for PubMedCentralID PMC10850404

• Secondary Cancer Risk in Breast Cancer with and without Radiotherapy: The Observational Health Data Sciences and Informatics (OHDSI) Cohort Study. Cancer research and treatment Kim, S., Boo, D., Yoo, S., Kim, B., Kim, K., Kim, K., Song, E., Kim, J., Ryoo, H. G., Paeng, J. C., Choi, I. Y., Ko, S., Yoo, I. R., Park, R. W., Lee, H. Y. 2025

  Abstract

  Radiotherapy is used to reduce the risk of breast cancer recurrence after surgery, but it is a potential cause of secondary cancer. We validated the risk of secondary cancer in primary breast cancer who received radiotherapy compared with those who did not from a matched cohort using a large-scale observational study of the Observational Health Data Sciences and Informatics (OHDSI) data network.A retrospective comparative cohort study using propensity score-matched cohorts was performed using two Observational Medical Outcome Partnership common data model databases, from tertiary general hospitals in South Korea. Among female patients who underwent surgery after the diagnosis of breast cancer, the risk of secondary primary malignant occurrence after 1:1 matching was analyzed.Among 27,078 patients with breast cancer, there was no significant difference in the risk of secondary cancer following radiotherapy in 4,268 patients after 1:1 propensity-score matching. Further, there were no significant differences in the sensitivity analysis according to age, latency period, and number of radiation treatments.There was no difference in the risk of secondary cancer in the patients diagnosed with breast cancer depending on whether or not radiotherapy was performed after surgery. In the future, it is necessary to analyze including data generated during cancer treatment.

  View details for DOI 10.4143/crt.2024.968

  View details for PubMedID 40506029

• Identifying potential medical aid beneficiaries using machine learning: A Korean Nationwide cohort study. International journal of medical informatics Kim, J., Park, S. H., Lee, H., Lee, S. K., Kim, J., Kim, S., Kwon, Y. J., Kim, K. 2025; 195: 105775

  Abstract

  To identify potential medical aid beneficiaries using demographic and medical history of individuals and analyzing important features qualitatively.This retrospective, national cohort, case-control study included data from the National Health Insurance Service (NHIS) in Korea between January 1, 2002 and December 31, 2019. Potential medical aid beneficiaries were classified using several machine learning models (linear models and tree-based models). Demographic data such as age, sex, region, insurance type, insurance fee, and medical history such as diagnosis, operation, statement, visits, and costs were collected. Those data were transformed into a one-dimensional vector for each individual, allowing machine learning models to learn. For feature importance calculation, we used the average gain across all splits for each feature.274,635 individuals were finally included in the study population, and 62,501 were classified as potential medical aid beneficiaries. XGBoost successfully classified potential medical aid beneficiaries with an AUROC of around 0.891. Assuming predicting before two years, the performance was still significant with an AUROC of around 0.832. Economic variables, such as insurance fees and several costs, turned out to be the most important, but variables regarding medical status, such as the results of blood tests and history of chronic diseases, were also important.Machine learning-based models successfully screened potential medical aid beneficiaries. Qualitative analysis of important features well reflected prior knowledge regarding public health. These findings can contribute to the soundness of healthcare finance and the improvement of public health.

  View details for DOI 10.1016/j.ijmedinf.2024.105775

  View details for PubMedID 39733535

• Prediction of reduced left ventricular ejection fraction using atrial fibrillation or flutter electrocardiograms: A machine-learning study. Digital health Kwon, S., Chung, S., Lee, S. R., Kim, K., Kim, J., Baek, D., Yang, H. L., Choi, E. K., Oh, S. 2025; 11: 20552076241311460

  Abstract

  Although the evaluation of left ventricular ejection fraction (LVEF) in patients with atrial fibrillation (AF) or atrial flutter (AFL) is crucial for appropriate medical management, the prediction of reduced LVEF (<50%) with AF/AFL electrocardiograms (ECGs) lacks evidence. This study aimed to investigate deep-learning approaches to predict reduced LVEF (<50%) in patients with AF/AFL ECGs and easily obtainable clinical information.Patients with 12-lead ECGs of AF/AFL and echocardiography were divided into those with LVEF <50% and ≥50%. A convolutional neural networks-based model customized to the study (AFibEFNet) and other deep-learning models were investigated. Electrocardiogram signals, ECG features, and clinical features (demographic information, comorbidities, blood cell counts, and blood test results) were collected for training. A hold-out test dataset was constructed using a different recruitment period. Five-fold cross-validation and calibration plots were used to evaluate performance.A total of 15,683 patients were analyzed (mean age, 70.0 ± 11.7 years; 61.2% men), with 82.2% having LVEF ≥50% and 17.8% having LVEF < 50%. Among the learning models, the AFibEFNet outperformed other models regarding area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and F1-score. Using ECG signals alone, the AFibEFNet model predicted reduced LVEF with AUROC of 0.798 (95% confidence interval [CI], 0.767-0.829) and AUPRC of 0.508 (95% CI, 0.434-0.564). For the AFibEFNet model, additional training with ECG and clinical features significantly improved AUROC (0.816 vs. 0.798, p = 0.04) and AUPRC (0.547 vs. 0.508, p < 0.001). The AFibEFNet model primarily focused on the R-wave, QRS onset and offset, and T-wave in ECG signals.Among the patients with AF/AFL, machine learning may predict reduced LVEF with 12-lead ECGs of AF/AFL.

  View details for DOI 10.1177/20552076241311460

  View details for PubMedID 39839953

  View details for PubMedCentralID PMC11748079

• Continual learning framework for a multicenter study with an application to electrocardiogram. BMC medical informatics and decision making Kim, J., Lim, M. H., Kim, K., Yoon, H. J. 2024; 24 (1): 67

  Abstract

  Deep learning has been increasingly utilized in the medical field and achieved many goals. Since the size of data dominates the performance of deep learning, several medical institutions are conducting joint research to obtain as much data as possible. However, sharing data is usually prohibited owing to the risk of privacy invasion. Federated learning is a reasonable idea to train distributed multicenter data without direct access; however, a central server to merge and distribute models is needed, which is expensive and hardly approved due to various legal regulations. This paper proposes a continual learning framework for a multicenter study, which does not require a central server and can prevent catastrophic forgetting of previously trained knowledge. The proposed framework contains the continual learning method selection process, assuming that a single method is not omnipotent for all involved datasets in a real-world setting and that there could be a proper method to be selected for specific data. We utilized the fake data based on a generative adversarial network to evaluate methods prospectively, not ex post facto. We used four independent electrocardiogram datasets for a multicenter study and trained the arrhythmia detection model. Our proposed framework was evaluated against supervised and federated learning methods, as well as finetuning approaches that do not include any regulation to preserve previous knowledge. Even without a central server and access to the past data, our framework achieved stable performance (AUROC 0.897) across all involved datasets, achieving comparable performance to federated learning (AUROC 0.901).

  View details for DOI 10.1186/s12911-024-02464-9

  View details for PubMedID 38448921

  View details for PubMedCentralID PMC11331660

• Risks of complicated acute appendicitis in patients with psychiatric disorders. BMC psychiatry Kim, J., Yang, C., Joo, H. J., Park, R. W., Kim, G. E., Kim, D., Choi, J., Lee, J. H., Kim, E., Park, S. C., Kim, K., Kim, I. B. 2022; 22 (1): 763

  Abstract

  Acute appendicitis often presents with vague abdominal pain, which fosters diagnostic challenges to clinicians regarding early detection and proper intervention. This is even more problematic with individuals with severe psychiatric disorders who have reduced sensitivity to pain due to long-term or excessive medication use or disturbed bodily sensation perceptions. This study aimed to determine whether psychiatric disorder, psychotropic prescription, and treatment compliance increase the risks of complicated acute appendicitis.The diagnosis records of acute appendicitis from four university hospitals in Korea were investigated from 2002 to 2020. A total of 47,500 acute appendicitis-affected participants were divided into groups with complicated and uncomplicated appendicitis to determine whether any of the groups had more cases of psychiatric disorder diagnoses. Further, the ratio of complicated compared to uncomplicated appendicitis in the mentally ill group was calculated regarding psychotropic dose, prescription duration, and treatment compliance.After adjusting for age and sex, presence of psychotic disorder (odds ratio [OR]: 1.951; 95% confidence interval [CI]: 1.218-3.125), and bipolar disorder (OR: 2.323; 95% CI: 1.194-4.520) was associated with a higher risk of having complicated appendicitis compared with absence of psychiatric disorders. Patients who are taking high-daily-dose antipsychotics, regardless of prescription duration, show high complicated appendicitis risks; High-dose antipsychotics for < 1 year (OR: 1.896, 95% CI: 1.077-3.338), high-dose antipsychotics for 1-5 years (OR: 1.930, 95% CI: 1.144-3.256). Poor psychiatric outpatient compliance was associated with a high risk of complicated appendicitis (OR: 1.664, 95% CI: 1.014-2.732).This study revealed a close relationship in the possibility of complicated appendicitis in patients with severe psychiatric disorders, including psychotic and bipolar disorders. The effect on complicated appendicitis was more remarkable by the psychiatric disease entity itself than by psychotropic prescription patterns. Good treatment compliance and regular visit may reduce the morbidity of complicated appendicitis in patients with psychiatric disorders.

  View details for DOI 10.1186/s12888-022-04428-7

  View details for PubMedID 36471298

  View details for PubMedCentralID PMC9721022

• Real-Time Evaluation of Cerebral Autoregulation Based on Near-Infrared Spectroscopy to Predict Clinical Outcome after Bypass Surgery in Moyamoya Disease. BioMed research international Kim, J., Ha, E. J., Kim, H. S., Park, E. Y., Lee, H. C., Choo, Y. H., Shim, Y., Kim, K., Kim, K., Lee, S. B. 2022; 2022: 3091660

  Abstract

  Impaired cerebral autoregulation (CA) can cause negative outcomes in neurological conditions. Real-time CA monitoring can predict and thereby help prevent postoperative complications for neurosurgery patients, especially those suffering from moyamoya disease (MMD). We applied the concept of moving average to the correlation between mean arterial blood pressure (MBP) and cerebral oxygen saturation (SCO2) to monitor CA in real time, revealing optimal window size for the moving average. The experiment was conducted with 68 surgical vital-sign records containing MBP and SCO2. To evaluate CA, the cerebral oximetry index (COx) and coherence obtained from transfer function analysis (TFA) were calculated and compared between patients with postoperative infarction and those who without. For real-time monitoring, the moving average was applied to COx and coherence to determine the differences between groups, and the optimal moving-average window size was identified. The average COx and coherence within the very-low-frequency (VLF) range (0.02-0.07 Hz) during the entire surgery were significantly different between the groups (COx: AUROC = 0.78, p = 0.003; coherence: AUROC = 0.69, p = 0.029). For the case of real-time monitoring, COx showed a reasonable performance (AUROC > 0.74) with moving-average window sizes larger than 30 minutes. Coherence showed an AUROC > 0.7 for time windows of up to 60 minutes; however, for windows larger than this threshold, the performance became unstable. With an appropriate window size, COx showed stable performance as a predictor of postoperative infarction in MMD patients.

  View details for DOI 10.1155/2022/3091660

  View details for PubMedID 37251497

  View details for PubMedCentralID PMC10212684

• Second primary malignancy risk in thyroid cancer and matched patients with and without radioiodine therapy analysis from the observational health data sciences and informatics. European journal of nuclear medicine and molecular imaging Kim, S., Bang, J. I., Boo, D., Kim, B., Choi, I. Y., Ko, S., Yoo, I. R., Kim, K., Kim, J., Joo, Y., Ryoo, H. G., Paeng, J. C., Park, J. M., Jang, W., Kim, B., Chung, Y., Yang, D., Yoo, S., Lee, H. Y. 2022; 49 (10): 3547-3556

  Abstract

  Risk of second primary malignancy (SPM) after radioiodine (RAI) therapy has been continuously debated. The aim of this study is to identify the risk of SPM in thyroid cancer (TC) patients with RAI compared with TC patients without RAI from matched cohort.Retrospective propensity-matched cohorts were constructed across 4 hospitals in South Korea via the Observational Health Data Science and Informatics (OHDSI), and electrical health records were converted to data of common data model. TC patients who received RAI therapy constituted the target group, whereas TC patients without RAI therapy constituted the comparative group with 1:1 propensity score matching. Hazard ratio (HR) by Cox proportional hazard model was used to estimate the risk of SPM, and meta-analysis was performed to pool the HRs.Among a total of 24,318 patients, 5,374 patients from each group were analyzed (mean age 48.9 and 49.2, women 79.4% and 79.5% for target and comparative group, respectively). All hazard ratios of SPM in TC patients with RAI therapy were ≤ 1 based on 95% confidence interval(CI) from full or subgroup analyses according to thyroid cancer stage, time-at-risk period, SPM subtype (hematologic or non-hematologic), and initial age (< 30 years or ≥ 30 years). The HR within the target group was not significantly higher (< 1) in patients who received over 3.7 GBq of I-131 compared with patients who received less than 3.7 GBq of I-131 based on 95% CI.There was no significant difference of the SPM risk between TC patients treated with I-131 and propensity-matched TC patients without I-131 therapy.

  View details for DOI 10.1007/s00259-022-05779-9

  View details for PubMedID 35362796

  View details for PubMedCentralID 4399282

• Transforming Thyroid Cancer Diagnosis and Staging Information from Unstructured Reports to the Observational Medical Outcome Partnership Common Data Model. Applied clinical informatics Yoo, S., Yoon, E., Boo, D., Kim, B., Kim, S., Paeng, J. C., Yoo, I. R., Choi, I. Y., Kim, K., Ryoo, H. G., Lee, S. J., Song, E., Joo, Y. H., Kim, J., Lee, H. Y. 2022; 13 (3): 521-531

  Abstract

  Cancer staging information is an essential component of cancer research. However, the information is primarily stored as either a full or semistructured free-text clinical document which is limiting the data use. By transforming the cancer-specific data to the Observational Medical Outcome Partnership Common Data Model (OMOP CDM), the information can contribute to establish multicenter observational cancer studies. To the best of our knowledge, there have been no studies on OMOP CDM transformation and natural language processing (NLP) for thyroid cancer to date.We aimed to demonstrate the applicability of the OMOP CDM oncology extension module for thyroid cancer diagnosis and cancer stage information by processing free-text medical reports.Thyroid cancer diagnosis and stage-related modifiers were extracted with rule-based NLP from 63,795 thyroid cancer pathology reports and 56,239 Iodine whole-body scan reports from three medical institutions in the Observational Health Data Sciences and Informatics data network. The data were converted into the OMOP CDM v6.0 according to the OMOP CDM oncology extension module. The cancer staging group was derived and populated using the transformed CDM data.The extracted thyroid cancer data were completely converted into the OMOP CDM. The distributions of histopathological types of thyroid cancer were approximately 95.3 to 98.8% of papillary carcinoma, 0.9 to 3.7% of follicular carcinoma, 0.04 to 0.54% of adenocarcinoma, 0.17 to 0.81% of medullary carcinoma, and 0 to 0.3% of anaplastic carcinoma. Regarding cancer staging, stage-I thyroid cancer accounted for 55 to 64% of the cases, while stage III accounted for 24 to 26% of the cases. Stage-II and -IV thyroid cancers were detected at a low rate of 2 to 6%.As a first study on OMOP CDM transformation and NLP for thyroid cancer, this study will help other institutions to standardize thyroid cancer-specific data for retrospective observational research and participate in multicenter studies.

  View details for DOI 10.1055/s-0042-1748144

  View details for PubMedID 35705182

  View details for PubMedCentralID PMC9200482