Leveraging EMR Data with AI to Forecast Genetic Testing Results in Neonatal Care using LUMINA

Accurate identification of patients likely to benefit from genetic testing in the neonatal intensive care unit (NICU) remains a clinical challenge. We introduce LUMINA (Learning from Unstructured Medical Information for Neonatal Assessment), a computational framework that leverages electronic medical record notes and Human Phenotype Ontology (HPO) embeddings to predict genetic testing outcomes. By integrating disease-specific databases, LUMINA captures phenotypic patterns that differentiate genetic from non-genetic patients. Evaluation of database-enriched HPO embeddings shows stronger disease-specific signals, while predictive models based on patient similarity matrices achieve robust discrimination and early identification of high-risk patients. Temporal analysis shows that LUMINA rapidly distinguishes genetic from non-genetic cases shortly after NICU admission, supporting timely decision-making. Importantly, the model’s predictions are interpretable: each patient’s classification is influenced by phenotypically similar patients, highlighting how shared clinical features drive the prediction. This approach provides a scalable, data-driven tool for neonatal precision medicine for future integration into clinical workflows.