Deterministic Graph-based Linkage of Brazilian Public Health Databases in Diffuse Large B-cell Lymphoma: a Reproducible Method for Public Health Management
Year: 2026
Type: article
Abstract: Introduction Fragmentation and anonymization of Brazilian Public Health Data (DATASUS) hinder the reconstruction of longitudinal patient trajectories, limiting healthcare service planning, care continuity and epidemiological research. Lacking a universal identifier acrossoutpatient (SIA), inpatient (SIH), and mortality (SIM) databases restrict cross-database analysis. Objectives To develop and validate a deterministic, graph-based linkage integrating SIA, SIH, and SIM to reconstruct diffuse large B-cell lymphoma (DLBCL) patients’ trajectories as a public health management tool. Material and Methods We included patients receiving outpatient DLBCL care (2015 and 2023). The pipeline uses intra-database graph connected components for deduplication into unique IDs. Each step applied multiple sequential blocking strategies combining fixed blocks (exact agreement) and dynamic blocks (controlled disagreement) usingbirth date, attendance/death date, facility, sex, race/color, ZIP code, municipality, pathological diagnosis date, and ICD-10. The SIA/SIH-SIM used a hierarchical cause-of-death prioritization: DLBCL (ICD-10 C83.3), diffuse non-Hodgkin lymphoma (ICD-10 C83.x), and other codes. As anonymization prevents direct performance measurement, the algorithm was validated with synthetic datasets with known true identity (calibrated on DATASUS) and with a semi-synthetic dataset fragmenting real encrypted National Healthcare Cards (CNS), evaluating performance through F1-score, precision, and recall. Clinical validation compared 5-year Kaplan-Meier survival of the linked São Paulo cohort with the São Paulo Hospital Cancer Registry (HCR-SP). Results From 47,335 distinct encrypted CNS in SIA, deduplication yielded 46,556 unique patients (1.6% reduction); 173,237 hospital authorizations collapsed into 85,644 patients. Among SIA patients, 18,037 (38.1%) had SIH records; of 21,551 eligible for SIM linkage, 5,998 (27.8%) had a death record. Synthetic validation showed a high-precision: SIA deduplication F1=0.999 (semi-synthetic 0.965), SIH deduplication F1=0.992, SIA-SIH precision 0.992/recall 0.788 (F1=0.878), SIA/SIH-SIM precision 0.771/recall 0.660 (F1=0.711). For São Paulo (2015–2019), 5-year survival was 58.7% (95% CI 56.8–60.6) versus 51.6% (49.1–54.2) from HCR-SP. Discussion The method reconstructs DLBCL trajectories from fragmented public data with high specificity; survival closely matched an independent cancer registry, supporting validity for real-world evidence. Conclusion Linkage performance was not uniform across systems, with the SIA/SIH-SIM integration showing lower precision and recall (F1=0.711), due to limited descriptive variables and restrictive blocking in SIM. Consequently, mortality outcomes should be interpreted with caution, as deaths may be missed or misattributed. These limitations, along with variable data quality and anonymization precluding direct performance measurement, were mitigated, though not eliminated, by synthetic and semi-synthetic validation. Despite these constraints, the graph-based deduplication and linkage successfully overcomes SIA-SIH fragmentation. By demonstrating high precision in integrating outpatient and inpatient records (F1=0.878), this linkage captures the complete longitudinal patient journey. Ultimately, this reproducible approach equips public health managers and researchers with a scalable, real-world evidence tool to monitor DLBCL care pathways, treatment burdens, and system-wide outcomes.
Language: English
Topic: Data Quality and Management
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7223158217