Neonatal respiratory diseases are a leading cause of morbidity and mortality in newborns, posing significant clinical challenges. Existing multimodal large language models (MLLMs) are largely trained on adult data, resulting in two major limitations: (1) a domain gap; (2) insufficient integration of multidimensional clinical context.
To address this, the authors collected two real-world clinical datasets, NeoCXR and NeoCXR‑EV, and introduced NeoRed—the first MLLM specifically designed for neonatal respiratory disease, filling the gap in neonatal diagnostic report generation.
To enable joint diagnosis from heterogeneous clinical information and chest X‑rays, they propose a Knowledge‑Logic‑Alignment (KLA) framework that constrains model behavior in three ways:
- Knowledge Prior Injection (KPI): injects neonatologist‑inspired diagnostic priors into multimodal representations, guiding the model to focus on disease‑relevant features across modalities;
- Diagnostic Logic Constraint (DLC): aligns the semantics of generated reports with multimodal diagnostic logic, ensuring that the reasoning process follows medical logic;
- Visual Semantic Alignment (VSA): establishes semantic correspondence between visual features and imaging conclusions, improving the utilization of chest‑film information.
Extensive experiments on NeoCXR show that NeoRed produces high‑quality diagnostic reports, achieving a ROUGE‑L of 53.29% and a Clinical Efficacy F1 of 65.19%, markedly outperforming existing MLLMs. The model also retains competitive performance on adult benchmarks (MIMIC‑CXR and IU‑Xray). The datasets will be made available upon request.
Review