Andre Massahiro Shimaoka

Introduction: Brazil's Unified Health System (SUS) established the country as a global reference in HIV/AIDS control, largely due to the timely implementation of universal access to combination antiretroviral therapy (cART) in 1996. However, recent policy setbacks—including service disruptions, testing shortages, and governance failures during the COVID-19 pandemic—highlight how institutional delays and neglect can yield severe health outcomes . Methods: In this retrospective ecological time-series study, we used counterfactual modeling to estimate the potential long-term consequences of delaying key public health interventions, simulating alternative cART implementation scenarios. Annual HIV/AIDS mortality data from 1980 to 2023 were obtained from Brazil's Mortality Information System (SIM/DATASUS) , and we employed the Prophet time-series framework to simulate delays of one year (1997) or two years (1998). Results : The historical introduction of cART in 1996 coincided with a marked decline in HIV/AIDS mortality. Counterfactual simulations indicated that a one-year delay could have resulted in 73,493 additional deaths by 2023, whereas a two-year delay could have resulted in 106,035 additional deaths. The effects were non-linear and extended well beyond the delay period, with mortality remaining elevated for decades. Conclusions : These findings demonstrate that delays and disruptions in HIV care generate persistent excess mortality. Counterfactual modeling underscores the critical need to safeguard institutional capacity, ensure continuity in testing and prevention, and prevent political or administrative setbacks in public health programmes.

OBJETIVO: Propor o framework Health Research Standard Process for Artificial Intelligence, adaptado da abordagem Cross-Industry Standard Process for Data Mining e aplicá-lo por meio de técnicas de séries temporais e inteligência artificial na análise dos dados do Sistema de Informação de Agravos de Notificação sobre violência contra mulheres em São Paulo, explorando dimensões sociodemográficas, espaciais e preditivas. MÉTODOS: A metodologia foi adaptada às pesquisas em saúde, substituindo a etapa original do Business Understanding por Research Understanding ao incorporar elementos técnico-científicos essenciais. Foram analisadas 80.148 notificações de violência contra mulheres entre 20 e 59 anos, residentes no município de São Paulo, no período de 2013 a 2023. As análises incluíram estatísticas descritivas, cálculo de razões de prevalência, decomposição e modelagem preditiva de séries temporais capturando tendências e sazonalidade, e análise geoespacial das notificações. RESULTADOS: Observou-se padrões temporais e características sociodemográficas das vítimas, com tendências crescentes nas notificações de violência física, psicológica e sexual após 2015 e comportamento sazonal. Entre as associações, identificou-se aumento de 32% na prevalência de violência sexual quando o agressor estava sob efeito de álcool e risco 2,47 vezes maior de violência sexual entre gestantes. A distribuição geoespacial revelou concentrações de notificações em áreas periféricas do município, como a zona sul, leste e central. A modelagem preditiva indicou que a tendência de aumento persistirá nos próximos 24 meses, com estimativas de taxas alcançando até 12 casos por 100.000 habitantes. CONCLUSÃO: A aplicabilidade do Health Research Standard Process for Artificial Intelligence foi eficaz como modelo para análises de dados e desenvolvimento de algoritmos preditivos em saúde pública. Foi possível propor uma matriz metodológica replicável para futuras pesquisas e intervenções baseadas em evidências.

Purpose: This study addresses Large Language Models (LLMs) pre-trained in Portuguese for healthcare applications, focusing on contextual embeddings. Research on LLMs for natural language processing (NLP) tasks in Portuguese is limited, especially within healthcare. Much of the existing research has focused on high-resource languages such as English. However, LLMs demonstrate potential in clinical decision support, diagnosis assistance, patient care, and other healthcare applications. In view of this, the present work assesses the current state of LLMs in Portuguese for healthcare. Methods: Our Systematic Literature Review (SLR) followed standard protocols: search, screening based on inclusion/exclusion criteria, quality assessment, data extraction, and analysis. Results: We identified 32 models, mostly based on BERTimbau, mBERT, and BioBERTpt. Adaptation strategies such as fine-tuning, domain-adaptive pre-training, training from scratch, and zero-shot learning have been the most prevalent. Several datasets have been used, including clinical records, social media, and scientific repositories. LLMs in Portuguese are being applied in mental health, general medicine, COVID-19, oncology, and other related areas, accomplishing classification tasks, followed by named entity recognition (NER), topic modeling, question answering, text generation, summarization, de-identification, and conversational agents. Conclusion: Our study identified key gaps and opportunities: (1) unexplored recent LLMs such as T5, Qwen, DeepSeek, BART, among others; (2) insufficient fine-tuning details, hindering reproducibility; (3) limited coverage of healthcare fields; (4) clinical and hospital data widely used but not shared; (5) social media data requiring caution due to potential inconsistencies; and (6) overlooked data privacy.

A Inteligência Artificial (IA) tem crescido globalmente e se consolidado como tecnologia estratégica em diferentes setores, ampliando a demanda por formação especializada. Nesse contexto, a pós-graduação lato sensu assume papel relevante na atualização profissional. Este estudo teve como objetivo analisar o perfil da oferta de cursos de especialização lato sensu relacionados à IA no Brasil. Trata-se de um estudo quantitativo e descritivo, baseado em dados do sistema e-MEC, no qual foram identificados 2.762 cursos. As análises consideraram a distribuição dos cursos segundo áreas do conhecimento, modalidade de oferta, carga horária, duração, número de vagas e localização regional. Os resultados evidenciam forte concentração da oferta nas áreas de Computação, embora a IA apresente caráter transversal e elevado potencial de aplicação em áreas como educação, saúde, negócios e engenharia. Observou-se predominância da modalidade a distância e concentração regional nas regiões Sudeste e Sul, além de elevada heterogeneidade nos modelos formativos, especialmente quanto à duração e ao número de vagas. Os achados indicam que a oferta de cursos ainda reproduz padrões históricos de concentração educacional. O estudo contribui ao oferecer um panorama nacional da formação em IA, fornecendo subsídios para o planejamento educacional e a formulação de políticas públicas.

A gestão estruturada de gastos pessoais e familiares tem ganhado relevância diante da crescente complexidade dos meios de pagamento e da necessidade de planejamento financeiro contínuo. No entanto, observa-se uma lacuna de estudos que descrevam de forma sistemática o desenvolvimento e a avaliação técnica de soluções abertas voltadas a esse contexto. Este estudo teve como objetivo descrever a concepção, o desenvolvimento e a avaliação técnica do MeuPatrimonio, um software livre e de código aberto destinado ao gerenciamento de gastos pessoais e patrimoniais. A pesquisa foi conduzida segundo a abordagem de Design Science Research, com delineamento descritivo e técnico, compreendendo as etapas de identificação do problema, definição do escopo, design e desenvolvimento do artefato, demonstração e avaliação. O sistema foi implementado com uma arquitetura web desacoplada baseada em Single Page Application integrada a uma API REST, utilizando Vue.js no frontend, Laravel no backend e PostgreSQL como sistema gerenciador de banco de dados. A avaliação do artefato foi realizada por inspeção técnica do código-fonte e da arquitetura, orientada por características selecionadas do modelo ISO/IEC 25010, com foco em adequação funcional, manutenibilidade e interoperabilidade. Os resultados indicam que o sistema implementa as funcionalidades definidas no escopo, apresenta organização estrutural compatível com a arquitetura proposta e possibilita integração por meio de interfaces padronizadas, com código-fonte disponibilizado para inspeção técnica, reuso e extensibilidade.

He growing availability of health data within the Brazilian Unified Health System (SUS) increases the potential for data-driven analysis. However, large datasets also introduce challenges related to volume, structure, and data integration. This study develops and evaluates an automated Extract, Transform, Load (ETL) pipeline for ingestion and preparation of data from the Ambulatory Information System (SIA-SUS). The architecture uses cloud computing infrastructure to support scalable processing. The study follows the Design Science Research approach, which focuses on the development and evaluation of technological artifacts. A pilot experiment processes data from January 2024 for three Brazilian states. The experiment includes approximately 3.2 million ambulatory records. Each execution runs five times to estimate operational variability. Results show stable pipeline performance across scenarios. The extraction stage accounts for the largest share of total execution time. Throughput remains relatively consistent despite differences in data volume. Linear regression between processed records and execution time produces a coefficient of determination of R² = 0.996. The result indicates an approximately linear relationship between data volume and processing time. The pipeline demonstrates operational feasibility and scalability potential. The architecture reduces the complexity of preparing large datasets from the SUS. The solution supports the development of analytical environments for public health data.

Analyze patterns of association between diagnoses related to domestic violence in hospital admissions and identify clinical-demographic profiles using unsupervised machine learning. Data from the SUS Hospital Information System between 2008 and 2023 were used, covering 90,798 hospitalizations of women aged 20 to 59 years with ICD-10 codes related to violence. Frameworks for data preparation were applied, as well as algorithms for identifying association rules between diagnoses and topic modeling. The hospitalization rate stabilized after 2012, with an average between 2.6 and 3.2 per 100,000 women. States such as São Paulo, Bahia, and Minas Gerais accounted for 46% of absolute cases; Rio Grande do Norte and Pará presented the highest proportional rates. The algorithm identified significant rules between types of injury and mechanisms of aggression. Latent Dirichlet Allocation modeling revealed nine distinct profiles, highlighting young women undergoing emergency surgeries due to polytrauma. The use of machine learning identified relevant clinical-epidemiological patterns to support prediction and surveillance strategies in Primary Care, contributing to addressing gender-based violence.

Resumo Analisar padrões de associação entre diagnósticos relacionados à violência doméstica em internações hospitalares e identificar perfis clínico-demográficos por aprendizado de máquina não supervisionado. Utilizaram-se dados do Sistema de Informações Hospitalares do SUS entre 2008 e 2023, abrangendo 90.798 internações de mulheres de 20 a 59 anos com códigos CID-10 relacionados à violência. Aplicaram-se frameworks para preparação dos dados, e algoritmos para identificação de regras de associação entre diagnósticos e modelagem de tópicos. A taxa de internações estabilizou após 2012, com média entre 2,6 e 3,2 por 100.000 mulheres. Estados como São Paulo, Bahia e Minas Gerais concentraram 46% dos casos absolutos; Rio Grande do Norte e Pará apresentaram as maiores taxas proporcionais. O algoritmo identificou regras significativas entre tipos de lesão e mecanismos de agressão. A modelagem Latent Dirichlet Allocation revelou nove perfis distintos, com destaque para mulheres jovens submetidas a cirurgias de urgência por politraumatismos. O uso de aprendizado de máquina identificou padrões clínico-epidemiológicos relevantes para subsidiar estratégias de predição e vigilância na Atenção Primária, contribuindo para o enfrentamento da violência de gênero.

Introduction: The COVID-19 pandemic, which began in early 2020, became a challenge for healthcare systems, especially in Brazil and the state of São Paulo. Non-pharmaceutical interventions, such as lockdowns and face masks, helped reduce transmission when combined with testing and tracing but were not sufficient alone to prevent high case numbers and deaths. On 5 May 2023, due to the sustained decline in severe cases and deaths, and an increase in global immunity through vaccination and prior infection, the World Health Organization (WHO) declared that COVID-19 no longer constituted a Public Health Emergency of International Concern (PHEIC). Consistent with the WHO's position, and the emergence of new SARS-CoV-2 variants continues to highlight the need for ongoing vaccination efforts. Additionally, post-acute sequelae of SARS-CoV-2 infection-commonly referred to as long COVID-have affected a significant portion of the infected population, adding to the burden on public health systems. Methods: Given this complex scenario, the present study investigates a comparative epidemiological analysis of COVID-19 in São Paulo state across three key phases: pre-vaccination, mass vaccination, and the post-PHEIC period. Results: Public data indicated that although immunization campaigns contributed to reduced cases, hospitalizations, and deaths, COVID-19 still caused more deaths than most infectious diseases during the analyzed period. The persistence of the pandemic occurred concurrently with declining vaccination coverage following the PHEIC declaration. Conclusion: To address this ongoing public health challenge, ongoing policies are needed, including continued sustained vaccination efforts, variant surveillance, public awareness campaigns, and long-term health system strengthening.

Timely identification of large-vessel occlusion (LVO) in ischemic stroke is essential for optimizing prehospital triage and enabling rapid mobilization of thrombectomy-capable teams. Traditional LVO screening tools are often lengthy and reliant on neurological examination skills that may be inaccessible to nonspecialists. To assess the ability of large language models (LLMs) to detect LVO using only free-text summaries, with or without National Institutes of Health Stroke Scale (NIHSS) data, in a national teleneurology service. We conducted a retrospective analysis of 2,887 suspected stroke cases across 21 spoke hospitals within a national TeleStroke network. Neurologist-authored case summaries were processed using natural language processing techniques, including text embedding and supervised machine learning classification. Contextual LLMs (BERTimbau, BioBERTpt, GPorTuguese-2) were evaluated with five algorithms. The Bootstrap method was employed to mitigate class imbalance, with performance averaging over 100 iterations. Of 1,060 cases included in the final dataset, 143 had confirmed proximal occlusions. Median Alberta Stroke Program Early CT Score (ASPECTS) was 9 and mean National Institutes of Health Stroke Scale (NIHSS) was 5.4 ± 2. AdaBoost paired with BioBERT yielded the highest accuracy (89.82%), precision (98.37%), and AUC (89.86%). Incorporating NIHSS as a numerical feature improved recall (87.60% with multilayer perceptron) and F1-score (89.05% with Dense Neural Network). BioBERT consistently outperformed other models, regardless of NIHSS inclusion. The LLM-based models demonstrated strong performance in identifying LVO using routine clinical narratives. These findings support the integration of NLP and ML in TeleStroke systems and underscore the need for further validation across larger, multilingual datasets to ensure generalizability and clinical applicability.