SUN-199 Searching Tree-based Models To Uncover Links Between Thyroid Disease And Respiratory Distress Syndrome

Abstract Disclosure: Y.D. Di Sarli: None. C.D. Juciara: None. K. Souza Santos: None. F.C. Guimarães: None. M.D. Motta: None. C.P. Camacho: None. Background: Thyroid hormones play essential roles in organogenesis, fetal glucocorticoid synthesis, and surfactant production. This study aimed to evaluate different decision tree strategies to investigate associations between maternal or fetal thyroid disease and respiratory distress syndrome (RDS). Methods: This study was approved by the ethics committee (CAAE: 44326821.8.3001.0086 and CAAE: 44326821.8.3002.5452). A cohort of newborns admitted to a tertiary public hospital's Semi-Intensive or Intensive Care Unit (SICU) was included. Maternal TSH levels, collected during pregnancy, were measured using the Advia Centaur XP platform. Newborn TSH was measured from filter-paper blood spots using the Abbott Architect i2000 immunoassay. Using Spyder software for Python, we constructed decision tree models with Grid Search, Random Forest, and XGBoost strategies to identify the best-performing models. Seventy percent of the newborns were randomly assigned to the training set and 30% to the test set. Overfitting was assessed by comparing training and test results. Results: Of the 1,031 newborns, 734 had complete clinical (gestational age, weight, first and fifth-minute Apgar scores, and thyroid disease history) and laboratory (maternal and neonatal) data for decision tree analysis. The cohort had a median gestational age of 37 weeks, a median birth weight of 2,569 g, a median first-minute Apgar score of 8, and a fifth-minute Apgar score of 9. Median maternal TSH was 1.81 mUI/L, and median newborn TSH was 3.1 mUI/L. Only 16 cases (2.18%) reported maternal thyroid disease, and 445 (60.6%) cases were diagnosed with RDS. In the training set, the initial decision tree achieved 70.8% accuracy, 63.7% sensitivity, and 82.7% specificity. However, in the test set, accuracy dropped to 65.6%, with 58.5% sensitivity and 74.5% specificity. The Random Forest model achieved higher performance in training (73.7% accuracy, 81.4% sensitivity, and 60.7% specificity) and slightly lower test performance (66.5% accuracy, 79.7% sensitivity, and 50.0% specificity). The XGBoost model demonstrated a possible overfitting, with training accuracy of 99.8%, specificity of 99.7%, and sensitivity of 100.0%, but poor test results (61.1% accuracy, 74.0% sensitivity, and 44.9% specificity). Conclusion: Tree-based decision strategies showed promising results. Notably, both maternal and newborn TSH levels appeared at different decision points in the initial tree, suggesting a potential influence of thyroid function on specific subgroups at risk for RDS. However, these findings must be interpreted with caution due to the inherent limitations of a retrospective study design. Presentation: Sunday, July 13, 2025