Sustainable Machine Learning Framework for Health Data Integration

The fragmentation and low integration of Brazilian public health databases constitute a critical barrier to evidence-based public policy evaluation. Information systems in the biomedical domain operate on pre-existing, autonomous, distributed, and heterogeneous databases, and the lack of interoperability hinders the tracking of therapeutic pathways, the measurementof health outcomes, and the identification of disparities across population groups. In parallel, machine learning pipelines for data integration still face high computational and energy costs, which are infeasible in the Brazilian public sector. To overcome this limitation, this pre-project proposes a machine learning and data science framework for health data integrationthat combines unsupervised learning via K-Means clustering and supervised learning via Random Forest, Naive Bayes, and a single-layer perceptron, with class-weight balancing and probability-based ensemble voting, thus enabling local execution with low energy consumption. The framework is designed with dataset-agnostic abstractions that decouple data sources fromalignment references, promoting interoperability across heterogeneous health systems. Its feasibility will be demonstrated through a proof of concept on the OAEI Anatomy track, with a public gold standard.