Fnu Afiahayati

A deep learning model with an imbalanced mammography dataset can bias models toward common benign BI-RADS categories and reduce recognition of less frequent malignant or high-risk categories. To address this issue, we propose B2M (Benign Share Benefit to Malignant), a model-agnostic framework for imbalance-aware multi-class BI-RADS classification in C-View mammography. B2M uses a two-phase training strategy that combines dual sampling with feature-space mixing. In Phase I, the model is trained with dual sampling, integrating instance-based and class-balanced sampling to increase minority-class representation while preserving majority-class diversity. In Phase II, the model is fine-tuned with feature-space mixing using samples from the two sampling streams. A soft-target regularization objective supervises the mixed features using labels from both streams, encouraging smoother decision boundaries across BI-RADS categories. We evaluated B2M on an imbalanced mammography cohort from the C-View EMBED dataset using stratified 5-fold cross-validation across multiple CNN backbones. C-View is a synthesized 2D mammographic image generated from 3D digital breast tomosynthesis data, capturing DBT-derived structural information while requiring less memory and computation than processing the full 3D image volume. Among these experiments, ResNeXt-50 with B2M achieved the highest balanced accuracy and Macro-F1 scores compared with the evaluated oversampling and mixing-based methods. This improvement requires an offline training-time overhead of approximately 2.81×, but it does not increase inference cost. Overall, the results suggest that B2M may be useful for imbalanced multi-class BI-RADS classification in C-View mammography. However, the findings are based on the EMBED cohort, and further validation, including external and prospective evaluation, is needed before clinical use.

The pretrained language model in Indonesian is already available for natural language processing tasks. However, this pre-trained model has been trained on Indonesian text, which has a different structure from the job description. Due to this, the pre-trained language model is less effective for skill recognition purposes. IndoBERTSkill is a novel pre-trained domain-specific language model that recognizes Indonesian language skills. It is built on the Bidirectional Encoder Representations from Transformers (BERT) architecture. IndoBERTSkill was trained on an extensive collection of Indonesian language texts from the Indonesian Wikipedia, the English Wikipedia, and Indonesian job descriptions from the job portal. IndoBERTSkill's performance was evaluated through two main approaches: (1) language modeling via Masked Language Model (MLM) prediction, and (2) fine-tuning on a custom annotated dataset (NERSkill) for Named Entity Recognition (NER) tasks. The fine-tuning process involved training a classification layer on top of the IndoBERTSkill model using BIO tagging to identify hard skills, soft skills, and technology entities. Similarly, the skill recognition model derived from IndoBERTSkill exhibits the highest F1-Score among various pre-trained language models, precisely at 87%, thus demonstrating robustness and strong generalizability for skill entity recognition in Indonesian job descriptions. IndoBERTSkill provides valuable resources for developing Indonesian natural language processing applications that require skills introduction. This could increase the accuracy and efficiency of skills recognition across various domains, including job matching, education, and training.

Asian genomic datasets possess unparalleled potential to advance global understanding of human genetic diversity. Encompassing the world’s largest population pool with diverse ethnicities, these datasets capture comprehensive genomic variations shaped by heterogeneous socioeconomic conditions, climate exposures, and clinical environments. However, current national genome initiatives across Asia demonstrate substantial disunity, stemming from limited cross-border communication and collaborative infrastructure, thereby diminishing their collective impact on biomedical research and precision medicine development. The MedHackathon Asia 2025 catalyzed crucial dialogues toward establishing a regional community dedicated to three pillars: harmonized biobank collaboration, standardized genomic data protocols, and cooperative governance frameworks. This multidisciplinary convening brought together researchers, clinicians, bioinformaticians, and national precision medicine program leaders from across Asia to share best practices, identify implementation challenges, and formulate foundational strategies for sustained cooperation. This community review synthesizes critical outcomes from these deliberations, emphasizing the imperative for continuous regional collaboration while advocating for the development of sustainable architectures enabling: (1) equitable biobank resource sharing, (2) genomic data standardization, and (3) ethical governance models. Through consolidation and expansion of this emerging network, Asian nations are expected to lead transformative contributions to global genomic science while ensuring appropriate representation in biomedical innovation. Such coordinated efforts promise to accelerate healthcare advancements with equitable benefits extending throughout the region and worldwide.

Skill entity recognition is a crucial task for aligning educational curricula with the evolving needs of the industry, particularly in multilingual job markets. This study introduces modified attention score transformer encoder (MAS-TENER), a novel transformer-based model designed to enhance the recognition of skill entities from Indonesian job descriptions. The proposed model modifies the attention mechanism by integrating relative positional embeddings and removing the scaling factor in self-attention. These improvements enhance the context of tokens, allowing for the accurate establishment of hard skills, soft skills, and technology skills. The MAS-TENER model was pre-trained and fine-tuned using a combinF.ation of job description datasets and additional corpora, achieving an F1-score of 90.46% at the entity level. The experimental results demonstrate the model's ability to handle unstructured, mixed-language job descriptions, with significant potential for curriculum reform and the development of new workforce capabilities. The study demonstrates the efficacy of the MAS-TENER model as an effective response for any natural language processing (NLP) task in low-resource languages. Moreover, the scope of long-term job market analytics in action research has been a key skill set in the education policy arena, demonstrating collaborative workforce capabilities.

Security incidents targeting supervisory control and data acquisition (SCADA) infrastructure are increasing, which can lead to disasters such as pipeline fires or even lost of lives. Man-in-the-middle (MITM) attacks represent a significant threat to the security and reliability of SCADA. Detecting MITM attacks on the Modbus SCADA networks is the objective of this work. In addition, this work introduces SMOTE tree-based autoencoder multi-stage detection (STAM) using the Electra dataset. This work proposes a four-stage approach involving data preprocessing, data balancing, an autoencoder, and tree classification for anomaly detection and multi-class classification. In terms of attack identification, the proposed model performs with highest precision, detection rate/recall, and F1 score. In particular, the model achieves an F1 score of 100% for anomaly detection and an F1 score of 99.37% for multi-class classification, which is preeminence to other models. Moreover, the enhanced performance of multi-class classification with STAM on minority attack classes (replay and read) has shown similar characteristics in features and a reduced number of misclassifications in these classes.

Blood disorders occur in one or several parts of the blood that affect the nature and function, and blood disorders can be acute or chronic. Blood disease consists of several types, such as anemia. Anemia is the most common hematologic disorder associated with a decrease in the number of red blood cells or hemoglobin, causing a decrease in the ability of the blood to carry oxygen throughout the body. Patients with anemia in Indonesia have increased for the age of 15-24 years. This study aimed to conduct a screening test for anemia using machine learning. It is expected to know the process of knowing the type of anemia suffered. The machine learning technique used to identify the cause of anemia is divided into four classes, namely Beta Thalassemia Trait, Iron Deficiency Anemia, Hemoglobin E, and Combination (Beta Thalassemia Trait and Iron Deficiency Anemia or Hemoglobin E and Iron Deficiency Anemia). This study would apply the K-Nearest Neighbor (KNN) and Random Forest (RF) methods to build a model on the data collected. The evaluation results using a confusion matrix in the form of accuracy, precision, recall, and f1-score against the KNN and RF methods are 79.36%, 59.40%, 62.80%, and 62.80%. In comparison, the RF is 87.30%, 90.89%, 78.40%, and 81.00%. From the results of comparing the two methods, the Graphic User Interface (GUI) implementation using python applies the RF method. The classifier that gets the highest value among all these parameters is called the best machine learning algorithm to perform screening tests for anemia.

The skin is an important part of the human body to be cared for. In healthy conditions, the skin has an important role in protecting the body from various disea The skin is an important part of the human body to be cared for. In healthy conditions, the skin has an important role in protecting the body from various disea