Machine learning solutions to identify and correct underdiagnosis of chronic non-communicable diseases in Brazil

• Grant number: 25/02245-8
• Support Opportunities: Regular Research Grants
• Start date: August 01, 2025
• End date: July 31, 2027
• Field of knowledge: Health Sciences - Collective Health - Epidemiology
• Principal Investigator: Alexandre Dias Porto Chiavegatto Filho
• Grantee: Alexandre Dias Porto Chiavegatto Filho
• Principal researcher abroad: Bruno Pereira Nunes
• Institution abroad: University of Illinois at Urbana-Champaign, United States
• Host Institution: Faculdade de Saúde Pública (FSP). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Associated researchers: Carine Savalli Redigolo ; Fabiano Novaes Barcellos Filho ; Gabriel Ferreira dos Santos Silva ; Marcela Quaresma Soares

Abstract

Non-Communicable Diseases (NCDs), such as diabetes mellitus, systemic arterial hypertension, and dyslipidemia, account for more than 80% of premature deaths globally and impose a high socioeconomic burden, particularly in low- and middle-income countries. In Brazil, it is estimated that approximately 45% of the adult population has at least one NCD. However, underdiagnosis rates exceeding 30% for hypertension and 50% for dyslipidemia and diabetes compromise the effectiveness of prevention and control policies. Underdiagnosis is largely driven by disparities in healthcare access, low adherence to screening protocols, and methodological limitations inherent to self-reported surveillance. This project proposes the development of machine learning models to estimate the actual prevalence of NCDs and identify patterns of unequal access to healthcare services in Brazil, using data from the 2013 and 2019 National Health Surveys (PNS).The methodological approach is based on predictive modeling using sociodemographic, clinical, and laboratory data. The response variable will be defined based on the discrepancy between self-reported diagnoses and clinical biomarkers, identifying underdiagnosed individuals. Supervised machine learning algorithms, including gradient boosting machines, random forests, and deep neural networks, will be employed, along with probabilistic models for prediction calibration. Transfer learning techniques will be used to adapt models trained in regions with greater data availability to contexts with lower sample representativeness, mitigating geographical and social biases.To ensure algorithmic fairness and mitigate preexisting biases, specific fairness metrics in predictive algorithms, such as equalized odds, demographic parity, and individual fairness, will be applied. Additionally, strategies such as resampling, class weighting, and post-processing adjustments will be implemented to ensure that predictions do not reinforce structural inequalities. Transparency and reproducibility will be ensured through complete model documentation, open-source code availability, and continuous algorithmic auditing.This project aligns with Brazil's Digital Health Strategy (2020-2028), enabling the integration of predictive models into SUS electronic systems to optimize NCD screening. The research also contributes to the Sustainable Development Goals (SDGs), particularly Goal 3 (Good Health and Well-being) and Goal 10 (Reduced Inequalities), as well as the Pan American Health Organization (PAHO) guidelines for strengthening public health systems.The developed models are expected to assist in the early identification of individuals at high risk of underdiagnosis, enabling targeted interventions and improving screening and triage strategies within SUS. Furthermore, the research findings could inform public policies aimed at reducing premature mortality associated with NCDs and optimizing healthcare resource allocation. The international collaboration between the University of São Paulo (USP) and the University of Illinois strengthens methodological innovation and fosters knowledge exchange, expanding the impact of this research in the fields of digital health and computational epidemiology. (AU)