Machine learning-based record linkage for brazilian health databases validated in Espírito Santo

A machine learning–based record linkage approach was developed within the TRAUMA project to support the integration of Brazilian Unified Health System (SUS) information systems. However, evidence comparing traditional linkage approaches with supervised machine learning models in large health datasets remains limited. This study evaluated the TRAUMA linkage method and compared it with CIDACS-RL, a widely used probabilistic linkage tool in Brazil. This methodological validation study used records from the Mortality Information System (SIM), Hospital Information System (SIH-SUS), Notifiable Diseases Information System (Sinan), and e-SUS Health Surveillance (e-SUS VS) from Espírito Santo, Brazil, between July 2018 and June 2025. Candidate pairs were generated through a multi-pass blocking strategy. The proposed method incorporated supervised machine learning to automate the classification of uncertain record pairs that would otherwise require manual review. A total of 106,363 labelled record pairs were used for model development and testing. Six machine learning algorithms were evaluated using sensitivity, specificity, precision, F1-score, and ROC-AUC. The best-performing algorithm was incorporated into the TRAUMA framework and compared with CIDACS-RL using a validated gold-standard dataset of 11,662 record pairs. LightGBM achieved the best performance. In the test dataset, the model obtained an ROC-AUC of 0.99996, sensitivity of 0.99731, precision of 0.99464, F1-score of 0.99597, and specificity of 0.99895. Performance remained stable in independent validation (ROC-AUC = 0.99988; F1-score = 0.99252). SHAP analysis identified date-of-birth variables and Jaro–Winkler and Levenshtein similarity measures as key predictors. Compared with CIDACS-RL, TRAUMA showed higher sensitivity, F1-score, and specificity estimates, with similar precision, and recovered more reference links. Both approaches demonstrated high performance for record linkage in administrative health databases. TRAUMA showed higher estimates of sensitivity, F1-score, and specificity while maintaining high precision and recovering more links contained in the reference dataset. These findings support both approaches as robust record linkage solutions, with the observed differences interpreted descriptively rather than as evidence of statistical superiority.