Unveiling the gap in heart failure: a Brazilian Unified Health System study.

Background: Heart failure (HF) decompensation is the leading cause of hospitalisations in developed countries and the third most common cause in Brazil. Underdiagnosis or misdiagnosis thereof remains a critical challenge and with significant implications. At the patient level, it delays appropriate treatment and disease management. At the health system level, it leads to inefficient population health strategies, distorted epidemiological estimates, and suboptimal resource allocation. Addressing this hidden HF population is therefore essential for improving disease surveillance, optimising care pathways, and supporting more effective and equitable health system planning in Brazil. In this sense, we aimed to quantify underreporting of HF in individuals in Brazil and to estimate their mortality trends. Methods: We identified potential HF patients using a guideline-based flowchart that combines ICD-10 codes (e.g. I21, I25, I42) and HF-related procedures (e.g. echocardiography, B-type Natriuretic Peptide (BNP) testing). Then, we developed a machine learning model for early prediction of HF using the Brazilian Department of Informatics of the Unified Health System (Departamento de Informática do Sistema Único de Saúde) mortality information system and ambulatory information system records, 2018-2022. A CatBoost algorithm was trained on a balanced cohort of patients (10 000 with HF and 9 995 without HF), restricting predictors to ≥12 months before the first I50 code to prevent leakage. Key predictors included age, chronic kidney disease, echocardiography, and lipoprotein disorders. We validated model performance in independent cohorts, including a low-prevalence cohort (~2% HF). Finally, we derived underreporting estimates via deterministic sensitivity analysis across scenarios from 0-100%. Results: The proxy identified approximately 54 000 potential HF cases/y that were unreported in ambulatory data, with deterministic analysis suggesting 12-41% underestimation. Mortality analysis showed that around 200 000 deaths could be linked to unreported HF. The CatBoost model achieved an area under the curve (AUC) of 0.91 (balanced cohort; accuracy = 0.82, recall = 0.81, F1 = 0.82) and maintained strong discrimination in low-prevalence settings (AUC = 0.84, sensitivity = 0.82), with excellent calibration (Brier = 0.124-0.136; ECE = 0.01-0.03). Conclusions: We found that HF underreporting in Brazil is substantial and carries significant mortality implications. Our ML model demonstrates high accuracy in early risk stratification using routine administrative data, aligning with clinical pathways. Implementing it as a screening tool could optimise resource allocation, but ethical considerations around false positives warrant careful deployment. Future work should focus on clinical validation and cost-effectiveness analysis within the Brazilian Unified Health System. Furthermore, our findings highlight significant implications for public health and clinical HF management, emphasising the necessity of strategies that promote early detection of HF and precise case recording. Addressing underestimation is crucial to optimise healthcare resources and improve patient outcomes. The results underscore the importance of accurate diagnosis and comprehensive management approaches for better HF case tracking. Further research should explore the public health impact of these underestimations in Brazil, particularly regarding its health system's financial resources.