Identificação da displasia broncopulmonar em recém-nascidos pré-termos de muito baixo peso por meio de biomarcadores urinários: aplicações não-invasivas com plataformas metabolômica e ATR-FTIR

Abstract

Introduction: Bronchopulmonary dysplasia (BPD) is a chronic lung disease affecting preterm infants, characterized by impaired alveolar and vascular development, resulting in compromised gas exchange and prolonged oxygen dependency. The high incidence of prematurity and the morbidity and mortality associated with BPD highlight the limitations of a diagnosis that remains largely based on late clinical outcomes, which fail to reflect the early immunoinflammatory alterations underlying the disease. This limitation hinders risk stratification and timely intervention, underscoring the need for the development of non-invasive biomarkers integrated with artificial intelligence to improve the prediction and management of BPD in very low birth weight preterm infants. Objective: To develop and evaluate non-invasive platforms based on urinary biomarkers integrated with artificial intelligence algorithms for the early prediction of BPD in very low birth weight preterm infants (VLBW-PTIs). Materials and Methods: A longitudinal study was conducted in a neonatal intensive care unit between January 2015 and September 2019, including preterm infants with gestational age <30 weeks and very low birth weight (<1,500 g). Infants were classified into two groups: BPD and non-BPD. Urine samples collected on the second day of life were analyzed using attenuated total reflectance Fourier-transform infrared spectroscopy (ATR-FTIR) and high-resolution mass spectrometry (MS) coupled with high-performance liquid chromatography (HPLC) for metabolite identification. Spectral data generated by both platforms were subsequently analyzed using artificial intelligence algorithms to assess diagnostic predictive performance (accuracy, sensitivity, and specificity). Statistical analyses included parametric and nonparametric tests for group comparisons, principal component analysis (PCA) of spectral data, machine learning-based classification algorithms, metabolomic analysis, univariable and multivariable logistic regression, internal bootstrap validation, receiver operating characteristic (ROC) curve analysis, and decision curve analysis. Results: In the ATR-FTIR analysis of 42 urine samples from very low birth weight (VLBW) preterm infants, including 20 with bronchopulmonary dysplasia (BPD) and 22 without BPD, the AdaBoost algorithm combined with second-derivative preprocessing demonstrated a sensitivity of 73%, a specificity of 81%, and an accuracy of 78% in classifying infants who did or did not develop BPD. The most relevant vibrational markers were associated with lipids, carbohydrates, and proteins. HPLC-MS metabolomic analysis of 88 urine samples from VLBW preterm infants, including 33 with BPD and 55 without BPD, identified two metabolites with significantly different abundances between the groups. The final multivariable model (M2), which combined sex, birth weight, gestational age, mechanical ventilation on the second day of life, and urinary Cys-Ser-Met, showed good discriminatory performance, yielding an AUC of 0.862 (bootstrap 95% CI, 0.775–0.937). At a cutoff of 0.347, the model achieved 80.7% sensitivity, 83.9% specificity, and 81.8% accuracy, with a positive predictive value of 90.2% and a negative predictive value of 70.3%. Conclusions: The results highlight the potential of a portable ATR-FTIR platform combined with artificial intelligence algorithms for the diagnostic screening of BPD using a non-invasive biological matrix such as urine. The urinary metabolomic approach identified Cys-Ser-Met and creatinine as differentially abundant metabolites in preterm infants with BPD, and the incorporation of Cys-Ser-Met into clinical variables improved model discrimination. However, these findings still require external validation in larger, multicenter cohorts. Together with our previous results obtained via ATR-FTIR, these data suggest that spectroscopic and metabolomic platforms can provide complementary biochemical signatures for the early identification of BPD, though their diagnostic and clinical utility must be prospectively confirmed prior to implementation in neonatal practice.