NeoRed Uses Knowledge-Logic Alignment for Neonatal Respiratory Diagnosis

Summary

Neonatal respiratory diseases are a major cause of neonatal morbidity and mortality, yet existing multimodal large language models face a domain gap because their training data is predominantly adult-focused. They also have difficulty integrating heterogeneous clinical context with chest X-rays for diagnosis. The study introduces NeoRed, which the authors describe as the first multimodal large language model tailored to neonatal respiratory disease and diagnostic report generation. It is trained and evaluated with two real-world clinical datasets, NeoCXR and NeoCXR-EV. Its Knowledge-Logic-Alignment framework combines three mechanisms: Knowledge Prior Injection, which adds neonatologist-inspired diagnostic priors to multimodal representations; Diagnostic Logic Constraint, which aligns generated reports with diagnostic logic; and Visual Semantic Alignment, which links visual features to imaging conclusions. On NeoCXR, NeoRed achieves a ROUGE-L score of 53.29% and a Clinical Efficacy F1 score of 65.19%, outperforming existing multimodal large language models in the reported experiments. It also retains competitive report-generation performance on the adult MIMIC-CXR and IU-Xray benchmarks. The datasets are planned to become available through an application process.