Diabetic foot ulcer (DFU) instance segmentation remains challenging due to severe class imbalance, boundary sensitivity, and limited labeled data. Building on our first-place DFUC 2024 winning system, we propose targeted improvements to enhance the Dice Similarity Coefficient (DSC). We used the original 2,000 image-mask pairs from DFUC 2024, augmenting 800 training images to obtain 4,000 effective training samples. Our improved model attaches a Mask R-CNN head to a DINOv2 backbone and adopts a strategy of partial fine-tuning. To better align optimization with the evaluation metric, we replace the default mask loss with a composite loss combining BCE and soft Dice. On the DFUC 2024 leaderboards, the DSC improved significantly: from 0.6719 to 0.6986 on the Live leaderboard and from 0.6519 to 0.6813 on the Validation leaderboard. Qualitative analyses confirm tighter alignment to ground truth and reduced symmetric-difference (XOR) error regions. These results demonstrate that principled, low-overhead adjustments to fine-tuning scope and loss design can yield measurable gains for DFU instance segmentation.

Multilingual medical speech recognition datasets are critical resources for building robust speech recognition systems that can reduce documentation workload and costs in healthcare while supporting clinical communication across diverse languages. In this study, we introduce the MultiMed Translation and Speech Synthesis (MultiMed-TSS) dataset, an extension of the original MultiMed dataset, where source languages were translated into multiple target languages in a many-to-many manner and paired with speech synthesized using a Text-toSpeech (TTS) model. The dataset comprises$\mathbf{1 6 5, 5 3 7}$speech-text pairs totaling 472 hours of audio across six languages: Chinese, English, French, German, Vietnamese, and Korean. To assess the effect of augmentation, we compared models fine-tuned with and without the TSS data. We observed that Vietnamese showed the largest gains while improvements in the other languages were modest. We have publicly released the MultiMed-TSS dataset at https://huggingface.co/datasets/SehwanMoon/MultiMed-TSS, to facilitate further research on medical speech recognition.

We propose an end-to-end Heart Rate Variability (HRV)-aware multi-task learning (MTL) framework for non-contact stress recognition from facial videos. Unlike approaches that rely on engineered rPPG features or separate preprocessing pipelines, our method jointly performs HRV regression and stress classification within a unified framework. Predicted HRV indices (SDNN, RMSSD, pNN50) are used as auxiliary signals to enhance both stress recognition accuracy and physiological interpretability. Experiments on the UBFC-Phys dataset demonstrate that our HRV-aware MTL model achieves 0.86 accuracy in 3-class classification and 0.94 in 2-class classification, comparable-to-better than a single-task baseline. While HRV regression alone yields only moderate correlation with ground truth, incorporating HRV as an auxiliary feature stabilizes classification and enhances physiological interpretability. These results highlight the effectiveness of combining physiological regression with psychological classification, paving the way for interpretable, non-contact stress assessment systems.

Recent advancements in large language models (LLMs) have opened new avenues in psychological counseling. This study leverages LLMs to develop chatbots capable of conducting empathetic and personalized counseling sessions by applying various prompt engineering techniques, including zero-shot, few-shot, meta-learning, Chain of Thought, and our newly developed Empathetic Meta-Chain (EMC) method. The EMC method demonstrated superior performance in empathy, response accuracy, interaction continuity, fluency, and understanding, as confirmed by expert evaluations. By integrating advanced empathetic strategies, the EMC chatbot significantly enhances its ability to support users' mental well-being through natural and engaging counseling interactions. These findings highlight the potential of LLM-based counseling chatbots to serve as effective tools in mental health support.

Classifying depressive symptoms using actigraphy data presents challenges, as highlighted by previous research. In this study, we constructed the TREND-P dataset, comprising 3,313 subjects from Chonnam National University Hospital (2021-2023), to analyze the impact of varying depressive symptom ratios. The dataset includes actigraphy, demographic, and HRV (Heart Rate Variability) data.We assessed the classification performance on two tasks: a challenging task differentiating subjects with PHQ-9 scores below 5 from those with scores of 5 or higher, and an easier task distinguishing subjects with PHQ-9 scores of 10 or higher from those below 5, excluding mild symptoms.Our results reveal that integrating multimodal data (actigraphy, HRV, and demographic data) slightly enhanced performance, particularly in easier tasks, with improvements. These findings underscore the potential of multimodal data integration in improving the classification of depressive symptoms. However, there are still performance limitations in the more challenging task of distinguishing between subjects with PHQ-9 scores below 5 and those with scores of 5 or higher.

This paper analyzes the Korean Disaster Survivor Interview (KDSI) Dataset, which was constructed as part of a disaster psychological recovery support service to train AI models for evaluating PTSD (Post-Traumatic Stress Disorder) symptoms among disaster survivors in South Korea. The KDSI dataset encompasses a wide range of information crucial for assessing PTSD symptoms and demonstrates its potential for AI model training. By demonstrating that the performance of AI models trained on well-known foreign PTSD datasets is significantly poor when evaluating PTSD in Korean disaster survivors, this study underscores the importance of using datasets that reflect the linguistic and cultural characteristics of the Korean population for accurate PTSD evaluation.

Diabetic foot ulcer (DFU) are a significant complication in diabetic patients, requiring accurate classification for effective treatment. This study investigates the use of self-supervised learning with Bootstrap Your Own Latent (BYOL) for feature extraction, combined with EfficientNetB0 as a downstream classifier. The proposed approach leverages unlabeled skin disease images for pretraining with BYOL, followed by fine-tuning on labeled data for four-class classification of infection, ischemia, both conditions, and other diabetic foot ulcers. We compared the performance of the BYOL pre-trained EfficientNetB0 model with a standard EfficientNetB0 classifier trained solely on labeled data. Our results demonstrate that the BYOL-based model achieved better performance in terms of accuracy, precision, recall, and f1-score. Additionally, Grad-CAM visualizations revealed that the BYOL-EfficientNetB0 model captures more accurate and relevant features compared to the baseline model. This study highlights the potential of self-supervised learning in improving the classification of DFU, especially in scenarios with limited labeled data.

Post-Traumatic Stress Disorder (PTSD) is characterized by various maladaptive responses that occur after experiencing or witnessing severe trauma, underscoring the importance of early detection and the development of automated diagnostic systems. This study aimed to construct the Korean Disaster Survivors Interview (KDSI) dataset and use it to develop and evaluate an AI-based diagnostic model for PTSD. We designed and trained four deep learning models for the classification of PTSD, anxiety, and depression using the KDSI dataset, with acoustic features extracted for performance analysis. The results demonstrated that the ensemble model (Model D) achieved the highest F1 score and AUC in anxiety classification and also showed strong performance in PTSD and depression classifications. Despite the potential noise and reliability issues associated with self-reported data collected online, the findings confirm that the KDSI dataset is a valuable resource for building AI models to assess disaster-related psychological conditions, including PTSD, anxiety, and depression.

Recently, there has been a significant amount of research conducted on classifying depression by extracting features from voice data. In this study, we enhanced the dataset by segmenting the voice data into 4-second intervals, followed by the extraction of voice features using MFCC (Mel Frequency Cepstral Coefficients) and Mel-spectrogram. After extracting these voice features, we compared the performance depending on the models and found that using MFCC features and classifying depression with the XGBoost model yielded the best performance. Looking forward, we aim to enhance the model’s performance by utilizing multimodal data, including voice, video, and text data.

Monitoring human behavior through wearable devices has potential in psychiatry. Among them, actigraphy data has been used to classify depression and detect depressive symptoms. We aim to collect a larger number of data to measure classification performance. This study evaluates the performance of classifying depressive symptoms solely on actigraphy data using both public (n=1549) and collected (n=3145) datasets. We found that there are challenges in classifying depressive symptoms from actigraphy data.