Domain-specific Data Research Articles

Large language models (LLMs) excel in many natural language processing tasks. However, their direct application to tabular, domain-specific clinical data remains challenging, as they lack innate mechanisms for reasoning over structured numerical features. This paper presents HealthAI-Prompt, a novel framework that systematically adapts LLMs for tabular clinical decision-making-specifically, predicting diabetes risk-through contextual prompts that combine detailed task descriptions with domain knowledge. Our domain knowledge integration leverages insights from high-performing machine learning models optimized via automated machine learning (AutoML) technique, together with local explanations for representative examples. These explanations, generated by multiple methods, are rigorously evaluated using fidelity, stability, and monotonicity metrics, ensuring reliability and clinical validity. The most accurate explanations are then embedded into the prompts, enabling the LLM to interpret structured features in a clinically meaningful way without fine-tuning. This approach uniquely bridges AutoML-driven predictive modeling with LLM-based reasoning over tabular inputs, improving predictive accuracy, and transparency in healthcare AI. Through comparative analysis of HealthAI-Prompt and various AutoML techniques under diverse data conditions, we offer insights into the impact of different prompt engineering strategies on model performance. As part of our evaluation, we also include Chain-of-Thought prompting as a baseline to contextualize the gains of our proposed method.

Self-supervised learning leads to improved performance in biparametric prostate MRI classification.

StoryPoint: GenAI-supported domain-specific data story authoring for enterprises

We will explore the current landscape of open-source research data management systems, focusing on platforms such as Invenio and Hyrax. Our workshop will enable users to interact with Invenio RDM and Samvera Hyrax data repositories. We will together walk through core features that are most frequently requested by researchers and research data administrators. These include: - Flexible importers and exporters to facilitate the large-scale ingestion and sharing of research data—both in terms of individual experiment size and volume of experiments. - Customizable workflows to support data review and publication processes. - Version control to track changes and maintain data integrity. - Granular authorization and authentication mechanisms to manage access rights. - Support for persistent identifiers - Metadata capture using a variety of metadata schemas. - Advanced search capabilities and the ability to view data directly within the system. We will also delve into emerging, nuanced features that are becoming increasingly important in modern research workflows, such as: - Offline data capture and seamless integration with the central data management system. - Support for archiving data not intended for publication, along with intuitive interfaces for managing such content. - Soft-delete functionality, where deleted data is moved to a temporary bin and later reviewed for either permanent archival or tombstoning. - Need for external data reviews, with strict access controls - enabled by repository-native support of Notify protocols and signposting, enhancing collaboration with external agencies and systems. Finally, we will discuss the shortcomings of current research data management systems, including: - Scalability challenges, such as limited support for diverse storage backends and declining performance with increased user and data volume. - Lack of versatile data viewers, especially for complex or domain-specific data types. - Insufficient emphasis on user experience (UX) across interfaces. - Limited integration of AI capabilities to assist in metadata extraction, data understanding, and intelligent presentation to end users. - A narrow focus on data storage and access, without sufficient support for data analysis and reproducibility. Current systems often manage only the data that has already been used for analysis, rather than assisting with the analytical process itself. Future systems will need to support rich metadata capture around data acquisition systems and analysis methods, and provide tools that facilitate replication and validation of research results.

Sustainable food development is crucial for minimizing environmental impacts and ensuring the capacity to provide sufficient food for both present and future generations. Many eco-friendly production methods have been adopted world-wide, including organic farming, regenerative agriculture, and plant-based alternatives, aimed at reducing greenhouse gas emissions, conserving water and soil resources, and promoting biodiversity. However, despite this development in sustainable production, consumer awareness and adoption of sustainable food choices remain limited, preventing full environmental and health impacts of these practices from being realized. This paper represents the design of an AI-powered chatbot, offering nutrition guidance, promoting sustainable and healthy daily food choices, while also addressing ethical considerations such as user privacy, fairness, and transparency in its design. The chatbot integrates artificial intelligence and large language models, adapted with domain-specific data on nutrition and sustainability, to engage users in conversations about healthy eating, food waste reduction, and eco-friendly diets. Its design combines a user-friendly interface, a curated knowledge base, and personalized recommendations informed by user preferences. Early evaluations suggest that the system can increase awareness and encourage more sustainable food choices. Ethical aspects such as privacy, transparency, and fairness are embedded in its development to promote responsible use of AI. Future enhancements may include integrating image-based calorie estimation to provide personalized nutritional feedback alongside sustainability guidance.

Recent advances in dynamic garment reconstruction boost virtual try-on with monocular video streams as inputs. However, existing literature has been intensively conducted on its sub-tasks, including static garment reconstruction and dynamic clothed human reconstruction, which are difficult to extend to dynamic garment reconstruction. The former bottleneck is mainly the lack of cross-frame correspondences and independent clothing topology on implicit garment fields, which results in the inability to obtain accurate motion information during dynamic clothing reconstruction and the absence of stable topology critical for downstream tasks, such as animation with physics engines. The latter usually binds the garment motion with body or skeleton motions, leading to rigid artifacts for loose-fitting garments. Our key idea is to build a diffusion based T-pose garments generator with a strong prior on garments structure. The garment generator is trained to generate 2D clothing representation, termed FOSUP, conditioned by a monocular video. FOSUP, defined as FOurier Spherical Unwrapping, enables a bidirectional mapping between FOSUPand the mesh through FFT and inverse FFT, which maintain spatial order and adjacency. Subsequently, this FOSUPis mapped back to 3D meshes through an inverse FFT process and transformed into pose space through a point transformation network to guide the three-dimensional reconstruction of the entire sequence. To sufficiently train our framework and address the lack of domain-specific data, we have constructed a large-scale garment MoCap dataset. This dataset captures the motion of various loose garments and includes multi-view raw images, frame-by-frame human motion annotations, raw scanned point clouds, topology-independent garment templates, and garment meshes with cross-frame correspondences. Comprehensive experiments have demonstrated that our unwrapping-based representation and diffusion-based framework significantly improve the performance and robustness of dynamic garment reconstruction.

The financial industry is a pillar industry of a country. Acting as an intermediary, finance provides direct or indirect capital services for various industries to meet diverse needs such as asset management, liability management, payment and settlement, and financial transaction processing. Compared with the digitalization of other industries, fintech is also characterized by massive transactions and a high degree of digitalization, giving AI numerous application scenarios and promising prospects in financial contexts. This paper investigates the practice of training large financial models based on general-purpose large models and industry data, their construction process, as well as their application cases and performance evaluation. It also analyzes and prospects the future scenarios and challenges of model development. It can be concluded that through training and fine-tuning with domain-specific financial data, large financial models can effectively enhance professional processing capabilitiesfor example, significantly improving the efficiency of test case generation and strengthening risk control capabilities. In the future, these models will develop towards lightweight architectures, with compressed parameter scales to meet the computational needs of small and medium-sized financial institutions, and intelligent agents will become an important application format. However, challenges such as data sharing difficulties, insufficient real-time model updates, and limited algorithm transparency still exist, requiring further exploration of secure data sharing mechanisms and dynamic model update technologies.

With the rapid development of artificial intelligence and machine learning technologies, automatic speech recognition (ASR) and text-to-speech (TTS) have become key components of the digital transformation of society. The Kazakh language, as a representative of the Turkic language family, remains a low-resource language with limited audio corpora, language models, and high-quality speech synthesis systems. This study provides a comprehensive analysis of existing speech recognition and synthesis models, emphasizing their applicability and adaptation to the Kazakh language. Special attention is given to linguistic and technical barriers, including the agglutinative structure, rich vowel system, and phonemic variability. Both open-source and commercial solutions were evaluated, including Whisper, GPT-4 Transcribe, ElevenLabs, OpenAI TTS, Voiser, KazakhTTS2, and TurkicTTS. Speech recognition systems were assessed using BLEU, WER, TER, chrF, and COMET, while speech synthesis was evaluated with MCD, PESQ, STOI, and DNSMOS, thus covering both lexical–semantic and acoustic–perceptual characteristics. The results demonstrate that, for speech-to-text (STT), the strongest performance was achieved by Soyle on domain-specific data (BLEU 74.93, WER 18.61), while Voiser showed balanced accuracy (WER 40.65–37.11, chrF 80.88–84.51) and GPT-4 Transcribe achieved robust semantic preservation (COMET up to 1.02). In contrast, Whisper performed weakest (WER 77.10, BLEU 13.22), requiring further adaptation for Kazakh. For text-to-speech (TTS), KazakhTTS2 delivered the most natural perceptual quality (DNSMOS 8.79–8.96), while OpenAI TTS achieved the best spectral accuracy (MCD 123.44–117.11, PESQ 1.14). TurkicTTS offered reliable intelligibility (STOI 0.15, PESQ 1.16), and ElevenLabs produced natural but less spectrally accurate speech.

This study introduces the Regularized Adaptive Weight Noise Injection-Based Evolutionary (RAWE) training approach, enhanced by generative Artificial Intelligence (AI), to optimize the dye recipe formulation in industrial textile manufacturing. RAWE integrates a self-adaptive evolutionary strategy with adaptive weight noise injection, dynamically balancing the exploration and exploitation during model training. A key innovation of RAWE is its use of generative AI to synthesize high-quality, domain-specific data, addressing the challenge of limited historical dyeing records. This synthetic data generation significantly improves the model generalization and robustness, enabling more accurate and reliable predictions in real-world industrial settings. The effectiveness of RAWE is demonstrated through its deployment in a real-world textile dyeing automation system, where it achieves significant improvements in dye recipe optimization. The results show that RAWE reduces the material waste, minimizes the production costs, and enhances the color consistency compared to traditional methods. By combining generative AI with adaptive evolutionary training, RAWE offers a scalable and practical solution for complex industrial processes, aligning with the latest advancements in automated Machine Learning (ML) and AI-driven optimization.

Vision foundation models, which have demonstrated significant potential in many multimedia applications, are often underutilized in the natural sciences. This is primarily due to mismatches between the nature of domain-specific scientific data and the typical training data used for foundation models, leading to distribution shifts. Scientific data often differ substantially in structure and characteristics, and researchers frequently face the challenge of optimizing model performance with limited labeled data of only a few hundred or thousand images. This work evaluates the performance of vision foundation models in astrophysics, with a focus on identifying the best practices for adapting these models to domain-specific datasets. We aim to establish a framework for selecting, fine-tuning, and optimizing these models for common tasks in optical and radio astronomy. We compared multiple foundation models, including self-supervised, weakly supervised, and distillation-based architectures, across two representative optical and radio datasets. Experiments involved different fine-tuning strategies, projector heads, and data preprocessing techniques, with performance evaluated on classification and detection metrics. Features extracted by specific foundation models improved classification accuracy for optical galaxy images compared to conventional supervised training. Similarly, these models achieved equivalent or superior performance in object detection tasks with radio images. However, classification performance for radio galaxy images was generally poor, often falling short of traditional supervised approaches. These findings suggest that selecting suitable vision foundation models for astrophysics applications requires careful consideration of the model characteristics and alignment with the specific requirements of the downstream tasks. This study demonstrates that vision foundation models can be effectively adapted to astrophysical applications, provided practitioners iterate on model selection, training strategies, and data handling. The proposed framework bridges the gap between these advanced models and the unique demands of astronomy, enabling broader adoption of deep learning in the field.

The increase in Arctic marine activity due to rapid warming and significant sea ice loss necessitates highly reliable, short-term sea ice forecasts to ensure maritime safety and operational efficiency. In this work, we present a novel data-driven approach for sea ice condition forecasting in the Gulf of Ob, leveraging sequences of radar images from Sentinel-1, weather observations, and GLORYS forecasts. Our approach integrates advanced video prediction models, originally developed for vision tasks, with domain-specific data preprocessing and augmentation techniques tailored to the unique challenges of Arctic sea ice dynamics. Central to our methodology is the use of uncertainty quantification to assess the reliability of predictions, ensuring robust decision-making in safety-critical applications. Furthermore, we propose a confidence-based model mixture mechanism that enhances forecast accuracy and model robustness, crucial for reliable operations in volatile Arctic environments. Our results demonstrate substantial improvements over baseline approaches, underscoring the importance of uncertainty quantification and specialized data handling for effective and safe operations and reliable forecasting.

The increasing regulatory scrutiny of social media, particularly regarding extremist content and misinformation, underscores the need for advanced threat detection systems. This article presents ALERT (Active Learning and Explainable AI for Robust Threat Detection in Telegram), a novel framework that enhances threat classification by introducing refined categories and creating tailored datasets. ALERT processes 2,301,110 replies from 17 Telegram channels, focusing on extreme content, with a dataset that predominantly reflects far-right discourse, consistent with activity trends on the platform. By leveraging an iterative active learning approach, it reduces labeling efforts by 86.5%, yielding a labeled dataset of 15,076 replies. ALERT’s RoBERTa+ model, pre-trained on domain-specific data, achieved over 90% in precision, recall, accuracy, and F1-score, demonstrating strong generalization for threat detection. The integrated Explainable AI (XAI) modules highlight key text features driving model predictions, ensuring transparency while maintaining performance. ALERT offers significant improvements in classification precision and user confidence, providing a critical tool for addressing digital threats while navigating regulatory and privacy challenges.

Accurately predicting the impact of point mutations on protein thermodynamic stability is essential for understanding structure-function relationships and guiding protein design. This challenge is particularly acute for transmembrane proteins (TMPs), which play vital roles in cellular signaling and drug targeting but remain underrepresented in structural databases. Existing predictors often rely on three-dimensional structures or multiple sequence alignments, limiting their applicability to TMPs due to poor structural coverage and alignment quality. Here, we present MEMO-Stab2, a fast and structure-independent deep learning framework for predicting mutation-induced stability changes in TMPs. MEMO-Stab2 reformulates the task as a binary classification problem, distinguishing destabilizing from neutral mutations based on a ΔΔG threshold of 0.4 kcal/mol. The model integrates multiview features within a Transformer-based architecture, utilizing embeddings from multiple pretrained protein language models (PLMs) and PLM-based structural predictions. By leveraging PLMs, it operates without requiring experimental 3D structures or explicit multiple sequence alignments, implicitly capturing both evolutionary and structural contexts from the amino acid sequence alone. Across internal and external transmembrane mutation data sets, MEMO-Stab2 consistently outperforms existing tools, including specialized predictors and a state-of-the-art general model even after it was fine-tuned on the same domain-specific data, achieving an F1 score of 0.92 on an internal benchmark. Further analyses confirm the model's robustness and specificity. It demonstrates strong generalization across diverse protein families with low sequence identity and shows superior performance in challenging biophysical contexts such as the transmembrane core and interfacial regions. Its validated computational efficiency enables large-scale mutation screening in minutes, offering a practical, robust, and powerful tool for transmembrane protein variant evaluation and engineering.

The intricate morphology of brain vessels poses significant challenges for automatic segmentation models, which usually focus on a single imaging modality. However, accurately treating brain-related conditions requires a comprehensive understanding of the cerebrovascular tree regardless of the specific acquisition procedure. Through image-to-image translation, our framework effectively segments brain arteries and veins in various datasets, while avoiding domain-specific model design and data harmonization between the source and the target domain. This is accomplished by employing disentanglement techniques to independently manipulate different image properties, allowing to move from one domain to the other in a label-preserving manner. Specifically, we focus on the manipulation of vessel appearances during adaptation, while preserving spatial information such as shapes and locations, which are crucial for correct segmentation. Our evaluation demonstrates efficacy in bridging large and varied domain gaps across different medical centers, image modalities, and vessel types. Additionally, we conduct ablation studies on the optimal number of required annotations and other architectural choices. The results obtained highlight the robustness and versatility of our framework, demonstrating the potential of domain adaptation methodologies to perform cerebrovascular image segmentation accurately in multiple scenarios.

Abstract The rapid advancement of large language models (LLMs) has opened up promising opportunities for their downstream applications in question-answering (QA), such as ChatGPT, ChatGLM, etc. However, such LLMs do not perform very well in domain-specific QA tasks without fine-tuning. But directly fine-tuning LLMs on domain-specific corpus data may lead to catastrophic forgetting, causing the LLMs to lose their general language capability. To address this problem, we propose the Knowledge-Enhanced Fine-Tuning (KEFT) method, an unsupervised fine-tuning approach to enhance the knowledge capability of LLMs in domain-specific QA tasks while preserving their general language capability. KEFT leverages the inherent language comprehension of pre-trained LLMs to generate synthetic-QA datasets from domain-specific corpus data autonomously for fine-tuning, and adopts a Low-Rank Adaptation (LoRA) method to further alleviate over-fitting. Furthermore, to enhance the representation of domain-specific knowledge, we introduce a knowledge-enhanced fine-tuning loss function, which encourages the model to learn the knowledge-question connection, thereby generating natural and knowledgeable answers. Our evaluations across multiple domain-specific datasets demonstrate that KEFT surpasses state-of-the-art fine-tuning approaches, enhancing the performance of various LLMs in QA tasks in both English and Chinese languages.

Fine-tuning large language models for interdisciplinary environmental challenges.

Optimizing Psychopharmacotherapy in Special Populations: Educating Language Models for Tailored Treatment Approaches

The integration of artificial intelligence, particularly large language models (LLMs), into veterinary education and practice presents promising opportunities, yet their performance in veterinary-specific contexts remains understudied. This research comparatively evaluated the performance of nine advanced LLMs (ChatGPT o1Pro, ChatGPT 4o, ChatGPT 4.5, Grok 3, Gemini 2, Copilot, DeepSeek R1, Qwen 2.5 Max, and Kimi 1.5) on 250 multiple-choice questions (MCQs) sourced from a veterinary undergraduate final qualifying examination. Questions spanned various species, clinical topics and reasoning stages, and included both text-based and image-based formats. ChatGPT o1Pro and ChatGPT 4.5 achieved the highest overall performance, with correct response rates of 90.4 and 90.8% respectively, demonstrating strong agreement with the gold standard across most categories, while Kimi 1.5 showed the lowest performance at 64.8%. Performance consistently declined with increased question difficulty and was generally lower for image-based than text-based questions. OpenAI models excelled in visual interpretation compared to previous studies. Disparities in performance were observed across specific clinical reasoning stages and veterinary subdomains, highlighting areas for targeted improvement. This study underscores the promising role of LLMs as supportive tools for quality assurance in veterinary assessment design and indicates key factors influencing their performance, including question difficulty, format, and domain-specific training data.

Reliability of LLM-based AI chatbots can be enhanced by using domain-specific models with a controlled knowledge base, especially relevant in the context of education and training of healthcare professionals and researchers. The aim was to produce an LLM data model with the goal of continuous education in radiation protection, that allows users to access reliable scientific information by querying it on specific topics, eliminating the need for manually perusing educational materials. A domain-specific LLM data model was developed and trained using custom knowledge from several domains, tested applying various test scenarios, and fine-tuned to ensure optimal selections of hyperparameters including top k, chunk size, temperature, max tokens etc. RESULTS: The final model produced reliable and accurate answers to a variety of users and queries based on controlled educational materials. Embedding model and similarity cutoff had the greatest impact on model performance. The developed model was trained and validated on radiation protection training material, allowing users to access information on topics including radiobiology and radiation protection in a quick and reliable manner.

Prostate cancer (PCa), one of the most common tumours in men, has a high death rate and is occasionally brought on by an inaccurate or delayed diagnosis. The urgent need for new accurate and dependable imaging-based diagnostic tools is highlighted by the low sensitivity and specificity of traditional screening methods such digital rectal examination and prostate-specific antigen (PSA) testing. In this paper, we provide a novel deep learning architecture for prostate cancer diagnosis based on the speed and accuracy of You Only Look Once, version 9 (YOLOv9). Our method minimizes the need for intensive post-processing by precisely localizing and classifying malignant lesions in multiparametric MRI scans via domain-specific data augmentation, adaptive anchor box adjustment, and fine-grained feature fusion. A carefully chosen and annotated collection of prostate pictures was divided into training, validation, and test sets using a patient-wise split to prevent data leakage. On the independent test cohort, the improved YOLOv9 model obtained a precision-recall value of 98%, an F1-score of 96%, a precision-recall of 91%, and a recall of 100%. These findings show a notable improvement in performance over the state-of-the-art techniques currently in use, especially when it comes to reducing false negatives, which is a crucial factor in clinical decision-making. Qualitative heat map analysis further showed that the model consistently focused on clinically important areas that were strongly aligned to expert radiologist observations. This attests to the model's interpretability and appropriateness for clinical real-time applications. In order to improve diagnostic confidence and facilitate early intervention, the suggested YOLOv9-based framework provides a quick, precise, and understandable method for PCa detection. Future studies will focus on merging multimodal imaging data and validating the model across bigger, multi-center surveys to ensure generalisability and therapeutic use.