DEFESA DE TESE DE DOUTORADO Nº 31

Aluno: Kayo Henrique de Carvalho Monteiro

Título: "IAra: A Platform to Combat Malaria in Brazil’s Legal Amazon"

Orientadora: Patricia Takako Endo

Coorientador: Vanderson Souza Sampaio (FMT)

Examinador Externo: Daniel Barros de Castro (FMT)

Examinador Externo: Antônio Alcirley da Silva Balieiro (FMT)

Examinador Externo: Domingos Sávio de Oliveira Santos Júnior (ITpS)

Examinador Interno: Eraylson Galdino da Silva

Data-hora: 31 de Julho de 2025 às 14h

Local: formato híbrido - UPE Caruaru

---

Resumo:

         "Malaria remains one of the most critical public health concerns in Brazil, with over 99\% of national cases occurring in the Legal Amazon region. The region’s geographic vastness, low population density, and limited healthcare infrastructure pose significant challenges for traditional epidemiological surveillance. To address this, the present study proposes the development of IAra, an intelligent platform for malaria incidence forecasting and epidemic classification, integrating spatial clustering and machine learning models within a unified and accessible web system. A robust and standardized methodology was employed throughout the study, encompassing data preprocessing, feature selection, spatial and temporal segmentation, model training, hyperparameter optimization, and epidemic classification. For predictive modeling, a comprehensive suite of techniques was used: classical statistical models (ARIMA), machine learning models (Support Vector Regression, Random Forest, XGBoost), and deep learning architectures (LSTM and GRU). This diverse modeling approach enabled comparative analyses across algorithms and scenarios, enhancing the methodological rigor and generalizability of findings. The first experimental study focused on the state of Amazonas, evaluating nine scenarios combining three types of malaria segmentation (All types, \textit{P. vivax}, and \textit{P. falciparum}) with three spatial grouping strategies (entire state, health regions, and k-means-based clusters). The results demonstrated that k-means clustering consistently improved predictive performance. For instance, in the case of \textit{P. vivax}, the ARIMA model achieved an RMSE of 0.027 and MAE of 0.020 when predicting for the whole state, whereas clustering reduced RMSE to as low as 0.016 and MAE to 0.013. For \textit{P. falciparum}, LSTM outperformed other models in certain clusters, revealing its strength in modeling localized, low-incidence dynamics. The second study extended the approach to classify predicted case counts into epidemic or non-epidemic events, using historical thresholds and rule-based criteria. By incorporating the duration and intensity of predicted outbreaks, the system supports not only numerical forecasting but also actionable alerts for health authorities. Results indicated that combining spatial clustering with classification techniques enables early identification of critical epidemiological shifts, especially in regions with volatile transmission patterns. To operationalize these results, IAra was implemented as a web-based platform integrating all methodological components. The system provides an intuitive and responsive interface that allows users to upload CSV datasets, view descriptive analytics, interact with heatmaps and time-series visualizations, and receive weekly forecasts and epidemic alerts. Built with modern technologies such as Next.js (React), Streamlit, and Plotly Dash, IAra ensures cross-device compatibility and real-time processing capabilities. The platform demonstrates strong potential for adoption by public health systems, offering both strategic and operational support for malaria surveillance in the Amazon. Furthermore, the methodological framework and system architecture are designed to be adaptable to other diseases and regional contexts, contributing to broader goals of data-driven, preventative health policy in alignment with the WHO’s malaria elimination strategies and the UN 2030 Agenda."