This paper introduces TLPSC, a novel clustering algorithm that combines Particle Swarm Clustering (PSC) with Prototype-Based Transfer Learning (PBTL). By leveraging knowledge transfer from a source domain, TLPSC enhances cluster quality, especially in scenarios with sparse or noisy data. Experimental results show TLPSC outperforms traditional clustering algorithms, such as K-Means, Gaussian Mixture Models (GMM), and the standard PSC, across some evaluation metrics, including NMI, ARI, Silhouette, and Davies-Bouldin. TLPSC shows its ability to maintain the data structure and generate cohesive clusters, proving to be a robust and efficient solution for clustering tasks.
Conversational agents in long-term healthcare often lack the ability to retain and recall information across sessions, leading to fragmented and impersonal interactions. This paper presents a scalable hybrid memory system that addresses this limitation by combining short-term memory for real-time dialogue with a persistent long-term memory layer based on MongoDB. The system, developed within the SALUS project, employs a dual-agent architecture: GenAI A handles immediate conversations using a short-term buffer, while GenAI B manages memory consolidation by generating and storing reflectionssemantic summaries of past interactions. Inspired by generative agent models, this design enables thematic memory organization, deferred storage, and keyword-based retrieval. Validated through use cases involving elderly patients and routine care scenarios, the architecture improves personalization, minimizes redundancy, and fosters trust in multi-session dialogues. The modular and extensible design supports future enhancements such as semantic search, explainable AI, and integration with socio-sanitary platforms like SALUS.
The healthcare sector is undergoing rapid transformation with the integration of Artificial Intelligence (AI), particularly Large Language Models (LLMs). These models, with their exceptional capabilities in natural language processing and generation, offer innovative solutions to address challenges in managing clinical workflows. Traditional Clinical Decision Support Systems often lack flexibility and struggle with real-time data integration, highlighting the need for more advanced approaches. This work proposes a clinical chatbot that leverages LLMs within a multi-agent framework to automate the prescription of medical exams and efficiently retrieve clinical data from patient records. Unlike diagnostic systems, the chatbot focuses on automating repetitive workflows, reducing clinician workload, and enhancing operational efficiency. The multi-agent architecture enables dynamic collaboration among specialized agents, while LLMs process natural language inputs, construct Fast Healthcare Interoperability Resources (FHIR) and Structured Query Language (SQL) queries, and generate actionable responses in real-time. This research underscores the transformative potential of LLMs, demonstrating their ability to streamline clinical workflows and improve resource management in healthcare environments.
This article presents a social simulator based on multi-agent systems designed to analyze the social well-being, health, and resource management of peasant families. Using an emotional BDI (eBDI) architecture, it explicitly integrates emotional factors influencing daily decision-making. Applied to vulnerable rural communities in Montes de María, Colombia, and developed using the Design Science methodology, the simulator integrates local expertise from agriculture, public health, and social domains. Validation through statistical and empirical evaluations confirmed the simulator’s accuracy in representing local dynamics. The results emphasize how efficient resource management and emotional well-being substantially enhance quality of life, illustrating its potential to inform public policies for sustainable rural development.
Academic dropout is one of the main challenges faced by higher education institutions, affecting both the efficiency of the educational system and the development of qualified professionals. These studies suggest that effective school inclusion, especially for students with special needs, requires understanding individual differences, collaborative and well-trained support among educators and families, and that simply adding technology or policies is not enough without addressing broader social and educational factors. This study proposes a machine learning-based predictive model to identify key factors associated with student dropout in the state of Pernambuco, Brazil, supporting data-driven decision-making in educational management. The dataset, sourced from the Faculdade Pernambucana de Saúde, includes academic, socioeconomic, and demographic variables. The modeling pipeline involved data normalization, feature selection using Boruta with Random Forest, class balancing via SMOTEENN, and hyperparameter tuning with GridSearchCV and stratified cross-validation. Multiple models were evaluated—Decision Tree, SVM, Logistic Regression, and XGBoost—using accuracy, F1-score, and ROC AUC. Additionally, SHAP was applied to interpret model outputs and assess feature contributions. Results highlighted the importance of academic performance, distance to campus, and financial status in predicting dropout, offering valuable insights for the development of more effective institutional retention policies.
