Geração de dados totalmente sintéticos para alta segura neonatal: uma metodologia orientada por conhecimento baseada em redes bayesianas

Abstract

Context: Neonatal hospital discharge is a critical stage in the transition of care, requiring multidimensional assessment of biophysiological, preventive, and healthcare-related conditions. Problem: Although Artificial Intelligence-based Clinical Decision Support Systems have the potential to reduce variability in professional judgment, their development in neonatology is limited by the scarcity of labeled records, legal restrictions on the use of sensitive health data, and the cold start problem. Motivation: The generation of fully synthetic data emerges as an alternative to support research, simulation, and computational development without direct exposure of real medical records. Justification: In sensitive domains such as neonatology, synthetic approaches guided only by data may depend on prior real datasets, reproduce biases, or generate clinically implausible combinations, making a knowledge-driven, interpretable, and auditable strategy relevant. Methodology: This dissertation proposes and evaluates a Bayesian Network-based methodology for generating fully synthetic data applied to decision support in safe neonatal discharge, without using real medical records in the modeling process. The methodology was developed incrementally, starting from an initial model with 11 clinical variables (Core11) and advancing to an expanded version with 18 variables (Core18). The causal network topology, Conditional Probability Tables (CPTs), normative weights, and clinical audit rules were defined based on specialized literature, normative documents, and clinical knowledge curation. The evaluation used internal validation protocols, including statistical fidelity, conditional inference, predictive utility under a Train on Synthetic, Test on Synthetic (TSTS) regime, SHAP-based explainability, threshold analysis, borderline case auditing, multivariate topological fidelity, and clinical edit checks. Results: The main experiments were conducted with 10,000 synthetic records in each model version. In Core11, the results indicated approval in protocols S1 and S2, a mean Kappa of 0.9445 in protocol S3, Hellinger Distance approved in 11/11 variables, and PCD-MAD of 0.0088 in protocol S7. In Core18, the expansion maintained good statistical, topological, and inferential fidelity, with S1 approved in 7/7 tests, S2 approved in 14/14 cases, a mean Kappa of 0.9209, Hellinger Distance approved in 18/18 variables, and PCD-MAD of 0.0097. However, relevant limitations were also observed, including divergence between explanatory importance and normative weights in S3SHAP, instability of decision thresholds in protocols S4 and S5, and the occurrence of clinical edit check violations in 36.9% of the synthetic records. Conclusions: It is concluded that the proposed approach is promising as a privacy-by-design synthetic data generation methodology for neonatal cold start scenarios, especially because it preserves probabilistic and topological properties of the model without relying on real patient data. At the same time, the findings show that global statistical coherence alone does not guarantee full individual clinical plausibility. Thus, the main contribution of this dissertation lies not only in generating synthetic neonatal data, but also in formulating an auditable pipeline for modeling, validation, and identification of methodological limits, indicating the need for future mechanisms of calibration, clinical filtering, conditional resampling, and expert validation.