Large Language Models (LLM) are becoming more essential in clinical text generation, where use of synthetic medical data is environmentally accurate and applicable for real-world healthcare applications. Existing LLMs often lack in specialized optimization and clarity, leading to incorrect outputs. These restrictions can make their references unreliable, particularly for sensitive clinical data. To overcome these problems, this research work suggests integrating generative adversarial networks with LLM to improve clinical data accuracy and reduce hallucinations. LLMs like LLaMA, BERT and GPT are broadly used in clinical settings for tasks such as summarizing patient notes and answering medical queries. Generative Adversarial Networks (GANs) are used to generate realistic synthetic clinical data, aiding privacy and data augmentation. The LDA model is added with GAN to identify the underlying topics in clinical documents, ensuring the synthetic text is coherent and thematically relevant. The use of Retrieval Augmented Generation (RAG) dynamically retrieves current medical knowledge and provides grounding responses with real-time evidence and minimizes outdated information. The first phase focuses on generating and validating synthetic clinical data using GANs and LDA to ensure high quality and domain alignment; the second phase focus on user interaction, where RAG retrieves relevant information in real time to answer queries, and an interactive interface enables seamless engagement and feedback. Continuous evaluation of NLP metrics demonstrates that the proposed Clinical Augmentation Generation and Retrieval Augmented Generation (CAG-RAG) framework outperforms the existing DALL-M approach in generating synthetic clinical text. For diagnosis-related data, the proposed CAG-RAG method achieves improvements of 15.7% in BLEU, 17% in ROUGE-1, and 17% in ROUGE-L scores. For medication-related data, the improvements were 20.8% in BLEU, 17.1% in ROUGE-1, and 17.25% in ROUGE-L. These results highlight the reliability, adaptability, and contextual accuracy for clinical applications.

The advent of the big data era has not only advanced the informatization of libraries but also opened unprecedented opportunities for their sustainable development. Libraries are no longer limited to traditional resource management; instead, they have embraced emerging technologies such as Web 2.0, mobile solutions, cloud computing, resource discovery systems, and big data platforms. While these developments provide a solid technological foundation, libraries must further enhance their ability to conduct data analysis, semantic processing, decision-making, and visualization in order to respond effectively to evolving user demands and complex information environments. This study contributes to that goal by discussing the application of multi-source data fusion in science and technology decision-making. It presents a comprehensive decision support framework that integrates semantic preprocessing techniques—including data cleaning, partition segmentation, and synonym merging—supported by Python’s Pandas library and Jieba’s text-cutting functions. Through this approach, the research successfully identified six science and technology text clusters and three mass technology-related clusters, thereby providing a refined view of user information needs and thematic structures within large-scale datasets. The findings demonstrate that a decision support framework based on multi-source data fusion can proactively detect and respond to user needs, moving libraries from passive service providers to active, intelligent participants in knowledge dissemination. This proactive transformation enriches the quality of information services, enables accurate and personalized decision support, and aligns with the demands of the new era defined by innovation-driven and intelligence-first strategies. Ultimately, this work highlights the value of integrating big data technologies into library management and decision-making systems. By bridging semantic analysis with multi-source data fusion, libraries can evolve into dynamic hubs of innovation, offering precise, context-aware services that not only enhance user satisfaction but also strengthen their role in supporting scientific research, technological advancement, and informed decision-making in the digital age.

Addressing the contemporary climate crisis is the need of the hour to protect both people and the planet. As countries embark on green energy revolution, focussing on achieving the United Nations (UN) 2030 agenda for Sustainable Development, guaranteeing universal access to affordable, reliable, and modern energy services stands out as an important goal. As part of the implementation of this goal, solar panel installation scheme has been undertaken by the government of India to encourage widespread adoption of green energy. This research work proposes an effective method to assess the acceptance of this scheme among users and the broader audience. User comments/ feedback from various social networking sites are analysed in this research work using Machine Learning techniques along with Explainable Artificial Intelligence (XAI) to make the machine learning models’ predictions more transparent. OpenAI Generative Pre-trained Transformer (GPT) language model is also used to automatically identify key implementation challenges of the scheme by creating a concise summary of the feedback shared by the users. This insight, based on the pain points of the users, can further help in providing recommendations and suggestions to appropriate stakeholders to improve the success rate of this scheme. Five machine learning models- Logistic Regression, Random Forest, Decision Tree, Extreme Gradient Boosting, and Stochastic Gradient Descent- were compared to choose the right technique for sentiment analysis. Among them, Logistic Regression and Stochastic Gradient Descent achieved an accuracy of 93% in predicting the sentiment. Our analysis showed around 63% of user feedback was positive indicating the public acceptance of green energy projects in India despite higher initial investments. The methodology and framework developed during this research work have immense reusability across similar government schemes (where transparency in sentiment analysis and sensitivity of public data are critical) in assessing their effectiveness and identifying areas where improvements are required.

Physical layer security (PLS) has been considered as a key technology to fulfill the information confidentiality request of modern and future communication networks. Therefore, diverse chaos-based wireless communication (CBWC) systems have been developed as low complexity and cost-effective PLS approaches when compared with the upper layer secrecy protocols. In particular, chaos-shift-keying (CSK) modulation schemes have attracted significant research efforts owing to the simple signal generation techniques and enhanced secrecy. However, the practical implementation of CSK for secure data transmission over realistic CBWC channels still needs further investigation. In this paper, the application of CSK based on multiple chaotic basis functions is examined over a band-limited CBWC channel with Rayleigh fading process. Lorenz and Chua chaotic oscillator circuits are used as basis signal functions for CSK modulation at the transmit side and chaos demodulation/synchronization at the receiver end. The impact of channel bandwidth and requisites of the front-end receiver is modeled as a low pass filter process. Performance results show that chaos filtering can greatly affects the physical features of employed signals at different levels. The achieved results confirmed that inadequate filter bandwidth can remarkably distort the state-space, signature waveform, and spectral components of CSK signals in disparate extents regardless of high SNR level. For target error rate and worst-case eavesdropping secrecy, this issue has a direct impact on decreasing the error security gap of CBWC system compared with the reference CSK schemes based on a single chaotic base function, even at a high received signal-to-noise ratio. As a feasible solution to mitigate the degradation in system reliability and secrecy, it is demonstrated that the designed filter bandwidth must include the effective spectral components of utilized chaotic signals.

In the era of rapid digital transformation, enterprise data security faces increasingly complex and dynamic threats. Traditional defense mechanisms are complicated to effectively respond to real-time risks, mainly when enterprises rely extensively on Internet of Things (IoT) devices. To address this problem, this paper proposes and implements a dynamic intelligent security assessment and early warning system based on ResNet-50 architecture and IoT technology. The system builds a distributed IoT data collection platform to collect multi-source data such as network traffic, device status changes, and user behavior in real time. It uses the optimized ResNet-50 model to analyze high-dimensional heterogeneous data streams accurately. The system is deployed in a cloud computing environment and can process large-scale data with low latency. It can instantly detect abnormal activities, conduct threat assessment, and issue alerts based on contextual information. Experimental results show that the system has an accuracy rate of 98.6% for distributed denial of service (DDoS) attacks and 96.2% for malware data leaks, with an average response time of 1.03 seconds, significantly better than traditional detection methods. This study provides an efficient and scalable solution for enterprise data security protection and lays a foundation for further integrating AI-driven models with IoT infrastructure.

Natural Language Understanding (NLU) is a branch of Natural Language Processing (NLP) that focuses on enabling computers to interpret human language with a level of understanding comparable to humans. NLU encompasses several tasks, including parsing sentences to understand grammatical structure, identifying word and phrase meanings, and determining user intent from natural language inputs. Many AI systems today—such as chatbots and virtual assistants—rely on NLU to accurately interpret and respond to user inputs in real time. This study addresses the challenge of accurately classifying user intents in multilingual intelligent virtual assistants a task critical for enhancing real-time human-computer interaction, by exploring the application of seven GPT-2 based models, leveraging their embedding matrices and tokenizers to design a robust intent-classification framework. The variation in the GPT-2 models in this study lies in the number of final layers and dimensional configurations used for classification. Through a large-scale case study with over one million utterances in 51 languages, the models were evaluated based on key metrics such as Accuracy, Precision, Recall, and F1-Score. Findings indicate that the GPT-256 model consistently achieved the highest values across these metrics, establishing it as the most accurate among the models tested. The GPT-256256 and GPT-128128 models followed closely, both of which showed competitive performance but with slightly lower accuracy than GPT-256. These results underscore the effectiveness of specific model configurations in improving NLU for virtual assistants, particularly in multilingual applications. The findings provide insight into optimizing AI systems for accurate goal classification, enhancing the ability of virtual assistants to understand and respond to diverse user inputs more precisely across languages, making them highly adaptable for global applications.

Diabetes, as a chronic condition affecting millions of people worldwide, requires early diagnosis and continuous monitoring to prevent complications. The rise of machine learning (ML) applications in healthcare offers promising approaches for diagnosing and managing diabetes more effectively. Machine learning models can analyse extensive amounts of data to identify patterns that may be invisible to human clinicians, improving diagnosis accuracy and enabling personalized care. This study investigates the performance of four machine learning models—Decision Tree, Logistic Regression, Random Forest, and Support Vector Machine (SVM)—in detecting diabetes using two types of data: traditional lab-based data and real-time accessed data from Internet of Things (IoT) sensors. Data was collected from continuous glucose monitors (CGMs) and wearables, as well as clinical lab records in Albania. The results revealed that machine learning models applied to IoT data significantly outperformed those applied to lab data, demonstrating higher accuracy and better predictive metrics. The continuous monitoring enabled by IoT devices allows for real-time detection of glucose fluctuations, providing earlier and more precise diabetes diagnosis. Additionally, integrating IoT with fog computing reduces latency and enhances on-time decision-making, allowing for prompt interventions in patient care. The study highlights the transformative potential of combining IoT, machine learning, and fog computing to revolutionize healthcare, particularly the management of chronic diseases such as diabetes. The findings suggest that IoT-based systems should be adopted to improve diabetes detection and monitoring, allowing for a shift toward proactive healthcare solutions. Future research could explore the application of these technologies for managing other chronic conditions and optimizing machine-learning models for large-scale datasets.

With the improvement of living standards and the enhancement of people's health awareness, participation in sports activities has received widespread attention. Traditional sports equipment, especially in educational environments, lacks the necessary technological advancements to provide precise guidance. The gap between this demand and available resources often leads to incorrect teaching methods, which may hurt students' sports training. To address these challenges, this paper proposes a motion action recognition system utilizing lightweight convolutional neural networks (CNN). This method effectively reduces the noise of sensor data, improves the accuracy and reliability of data, and lays a solid foundation for model training through one-dimensional median filtering and Z-score standardization. Optimize the CNN architecture by adjusting key parameters such as network structure, convolution kernel size, and convolution stride, which are fine-tuned based on training data to maximize the model's recognition ability. The research results provide valuable insights into the effectiveness of teaching techniques and targeted feedback for improving sports training. After sufficient training, the system performed excellently on test data, accurately identifying erroneous movements across various sports actions, particularly in critical areas such as stroke movements, with an accuracy rate of up to 97.82% and an RMSE as low as 1.71%. These results demonstrate the model's high precision and robustness. The system has shown great potential in addressing the current shortage of professional coaches by providing automatic, real-time feedback on motion accuracy.

Ensuring semantic interoperability between heterogeneous systems remains a challenging task due to the structural complexity and diversity of ontological representations. Traditional ontology alignment methods often focus on local features, overlooking important semantic relationships beyond direct neighbourhoods. Here, we introduce CNPMap, a novel alignment approach that addresses this limitation by capturing non-local semantic similarities using a critical node-based partitioning strategy. CNPMap operates in three stages. First, it generates an initial alignment using a hybrid linguistic similarity measure. Then, a graph-based partitioning method exploits the Critical Node Detection Problem (CNDP) to divide ontologies into semantically coherent components. Finally, a context-aware similarity enhancement phase refines the alignments using a sigmoid function that modulates similarities based on both partition-level and entity-level relationships. We evaluated CNPMap on the OAEI 2023 Conference track. The approach improved the F-measure on several ontology pairs by 3% to 6% compared to baseline lexical matchers. For instance, the F-measure increased from 0.69 to 0.74 on the cmt–conference pair and from 0.76 to 0.82 on the cmt–sigkdd pair. CNPMap also achieved a precision of 0.75, outperforming most participating systems. However, its recall was slightly lower due to the conservative threshold used during the initial alignment phase. Our study reveals that integrating partition-based context into similarity computation significantly improves alignment quality, especially for complex ontologies. Future enhancements will focus on improving recall through adaptive thresholds and learning-based parameter tuning.