BSBeomgi So

Beomgi So

*This is an algorithmically generated author record based on publicly available information.

Beomgi So has a total of 30 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Optimizing Small Local Language Models for Culturally Competent Mental Health Counseling: Comparative Evaluation With GPT-4o by Psychiatrists

    by Jaewook Han, Beomgi So, Whanbo Jung, Minwoo Kim, Hoon Kim, Daun Shin

    Background: Large language models (LLMs) show promise for mental health counseling applications. However, their deployment faces significant challenges, including data privacy concerns requiring on-premise solutions, high operational costs, and Western-centric biases that limit cultural appropriateness for non–English-speaking populations. Objective: This study aimed to develop and evaluate culturally competent small language models (SLMs) with approximately 7 to 8 billion parameters for Korean mental health counseling and to validate a 2-stage hybrid evaluation framework combining LLM-as-a-Judge screening with expert clinical validation. Methods: We constructed a dataset of 1558 diary-based counseling interactions from a Korean mental health mobile app (Maeumpyeonuijeom). Three Korean-optimized base models (SKT A.X 4.0 Light {7B}, LG EXAONE 3.5 {7.8B}, and Meta Llama 3.1 {8B}) were fine-tuned using an aggressive fine-tuning strategy with high-rank low-rank adaptation (r=256) over 20 epochs. Inference parameters were optimized using GPT-5 (gpt-5-2025-08-07) as an LLM-as-a-Judge. The optimized SLMs and GPT-4o (gpt-4o-2024-08-06) were evaluated by 10 board-certified psychiatrists (mean clinical experience of approximately 10 years) on 50 counseling scenarios across 4 dimensions (individualization, supportiveness, cultural appropriateness, and safety) using 5-point Likert scales. Wilcoxon signed-rank tests with Holm correction were used for statistical comparison, and intraclass correlation coefficients (ICCs) assessed interrater reliability. Results: No statistically significant difference from GPT-4o was detected for SKT A.X 4.0 Light in supportiveness (mean 3.39, SD 0.74 vs mean 3.45, SD 0.90; P=.26) and cultural appropriateness (mean 3.30, SD 0.77 vs mean 3.29, SD 0.79; P=.58). SKT A.X 4.0 Light ranked first in cultural appropriateness among all models. Interrater reliability analysis revealed higher agreement among psychiatrists for local SLMs (ICC=0.575 for SKT A.X 4.0 Light in cultural appropriateness) compared with GPT-4o (ICC=0.211). GPT-4o maintained significantly higher safety scores than all SLMs (P<.001). The LLM-as-a-Judge demonstrated moderate-to-good correlation with human evaluation but exhibited a ceiling effect in safety scoring. Conclusions: A 7 to 8-billion-parameter SLM fine-tuned on culturally specific data can achieve performance comparable to GPT-4o in supportiveness and cultural appropriateness for mental health counseling. These findings support the feasibility of locally deployable, privacy-preserving mental health AI systems based on culturally adapted SLMs. However, the safety gap necessitates human expert oversight, and success depends on the availability of high-quality language-specific base models.

  • Open Access

    Predicting Probable Persistent PTSD Following the Sewol Ferry Disaster: Development of an AI Algorithm Based on Psychological Assessments and Biological Markers

    by Daun Shin, Beomgi So, Jeong-Ho Chae, Byung-Joo Ham, So Hee Lee

    Objective Prolonged post-traumatic stress disorder (PTSD) carries substantial personal and public-health costs, yet early identification of individuals at risk remains difficult. This study aimed to develop a novel artificial intelligence (AI)-based predictive algorithm using psychological, biological, and psychosocial data collected at initial assessment to enhance early identification of individuals at heightened risk for probable persistent PTSD.Methods This study included 88 bereaved family members of the victims of the Sewol ferry disaster, divided into probable persistent PTSD (n=67) and probable remitted (n=21) groups based on 4-year follow-up assessments. Demographic, blood test, and psychological data were collected during initial evaluations. Models compared linear discriminant analysis (LDA) with tree-based learners under stratified cross-validation; class imbalance was addressed with Borderline Synthetic Minority Oversampling Technique, and recursive feature elimination identified parsimonious predictors.Results The LDA model demonstrated the highest performance, with an area under the receiver operating characteristic curve of 0.858. Psychosocial features dominated prediction: higher anxiety, depression, insomnia, and intrusive rumination increased risk, whereas greater positive resources and functional social support were protective. Routine physiological markers showed limited incremental value.Conclusion Findings support a practical intake pathway in which brief psychosocial measures are used with an interpretable classifier to triage high-risk individuals to targeted interventions. External validation, calibration, decision-curve analysis, and broader biomarker panels are needed to confirm transportability and optimize clinical utility.

  • Open Access

    Optimizing Small Local Language Models for Culturally Competent Mental Health Counseling: Comparative Evaluation With GPT-4o by Psychiatrists (Preprint)

    by Jaewook Han, Beomgi So, Whanbo Jung, Minwoo Kim, Hoon Kim, Daun Shin

    BACKGROUND Large language models (LLMs) show promise for mental health counseling applications. However, their deployment faces significant challenges, including data privacy concerns requiring on-premise solutions, high operational costs, and Western-centric biases that limit cultural appropriateness for non–English-speaking populations. OBJECTIVE This study aimed to develop and evaluate culturally competent small language models (SLMs) with approximately 7 to 8 billion parameters for Korean mental health counseling and to validate a 2-stage hybrid evaluation framework combining LLM-as-a-Judge screening with expert clinical validation. METHODS We constructed a dataset of 1558 diary-based counseling interactions from a Korean mental health mobile app (Maeumpyeonuijeom). Three Korean-optimized base models (SKT A.X 4.0 Light {7B}, LG EXAONE 3.5 {7.8B}, and Meta Llama 3.1 {8B}) were fine-tuned using an aggressive fine-tuning strategy with high-rank low-rank adaptation (r=256) over 20 epochs. Inference parameters were optimized using GPT-5 (gpt-5-2025-08-07) as an LLM-as-a-Judge. The optimized SLMs and GPT-4o (gpt-4o-2024-08-06) were evaluated by 10 board-certified psychiatrists (mean clinical experience of approximately 10 years) on 50 counseling scenarios across 4 dimensions (individualization, supportiveness, cultural appropriateness, and safety) using 5-point Likert scales. Wilcoxon signed-rank tests with Holm correction were used for statistical comparison, and intraclass correlation coefficients (ICCs) assessed interrater reliability. RESULTS No statistically significant difference from GPT-4o was detected for SKT A.X 4.0 Light in supportiveness (mean 3.39, SD 0.74 vs mean 3.45, SD 0.90; P=.26) and cultural appropriateness (mean 3.30, SD 0.77 vs mean 3.29, SD 0.79; P=.58). SKT A.X 4.0 Light ranked first in cultural appropriateness among all models. Interrater reliability analysis revealed higher agreement among psychiatrists for local SLMs (ICC=0.575 for SKT A.X 4.0 Light in cultural appropriateness) compared with GPT-4o (ICC=0.211). GPT-4o maintained significantly higher safety scores than all SLMs (P<.001). The LLM-as-a-Judge demonstrated moderate-to-good correlation with human evaluation but exhibited a ceiling effect in safety scoring. CONCLUSIONS A 7 to 8-billion-parameter SLM fine-tuned on culturally specific data can achieve performance comparable to GPT-4o in supportiveness and cultural appropriateness for mental health counseling. These findings support the feasibility of locally deployable, privacy-preserving mental health AI systems based on culturally adapted SLMs. However, the safety gap necessitates human expert oversight, and success depends on the availability of high-quality language-specific base models.

  • Open Access

    An artificial intelligence model integrating clinico-laboratory data and single nucleotide polymorphism-based genomic risk for prostate cancer diagnosis in Korean men

    by Jae Hung Jung, Gong Ho Han, Beomgi So, Myunghee Hong, Si Hoon Ahn, Sung Hyun Lim, Soo Min Han, Hongzoo Park, Sang Wook Lee, Geehyun Song, et al.

    Jung JH, et al. Investig Clin Urol. 2026 ;67:e35. https://doi.org/10.4111/icu.20260002

  • Open Access

    A Machine Learning Model for Prostate Cancer Prediction in Korean Men

    by Sukjung Choi, Beomgi So, Shane Oh, Hongzoo Park, Sang Wook Lee, Geehyun Song, Jong Min Lee, Jung Ki Jo, Seon Hyeok Kim, Si Eun Lee, et al.

    Purpose: Unnecessary prostate biopsies for detecting prostate cancer (PCa) should be minimized. Therefore, this study developed a machine learning (ML) model to predict PCa in Korean men and evaluated its usability.Materials and Methods: We retrospectively analyzed clinical data from 928 patients who underwent prostate biopsies at Kangwon National University Hospital between May 2013 and May 2023. Of these, 377 (41.6%) were diagnosed with PCa, and 551 (59.4%) did not have cancer. For external validation, clinical data from 385 patients aged 48–89 years who underwent prostate biopsies from September 2005 to September 2023 at Wonju Severance Christian Hospital were also included. Twenty-two clinical features were used to develop an ML model to predict PCa. Features were selected based on their contributions to model performance, leading to the inclusion of 15 features. A meta-learner was constructed using logistic regression to predict the probability of PCa, and the classifier was trained and validated on randomly extracted training and test sets at an 8:2 ratio.Results: The prostate health index, prostate volume, age, nodule on digital rectal examination, and prostate-specific antigen were the top 5 features for predicting PCa. The area under the receiver operating characteristic curve (AUC) of the meta-learner logistic regression model was 0.89, and the accuracy, sensitivity, and specificity were 0.828, 0.711, and 0.909, respectively. Our model also showed excellent prediction performance for high-grade PCa, with a Gleason score of 7 or higher and an AUC of 0.903. Furthermore, we evaluated the performance of the model using external cohort clinical data and achieved an AUC of 0.863.Conclusions: Our ML model excelled in predicting PCa, specifically clinically significant PCa. Although extensive cross-validation in other clinical cohorts is needed, this ML model is a promising option for future diagnostics.