Abstract: Predicting small for gestational age (SGA) and large for gestational age (LGA) newborns is crucial for preventing adverse pregnancy outcomes and improving neonatal health. Existing purely data-driven methods have achieved promising performance in SGA-LGA newborn prediction, but most of them lack model interpretability, raising doubts in clinical auxiliary diagnosis and failing to provide targeted interventions. To address this challenge, this paper proposes an obstetric knowledge-driven large language model, termed Obstet-LLM. Specifically, Obstet-LLM is pre-trained on the content from professional knowledge bases in the field of obstetrics. This pre-training enables the model to understand and integrate domain-specific knowledge and clinical insights. Subsequently, the model undergoes prompt engineering and instruction fine-tuning to learn precise predictions of neonatal growth outcomes. To enhance model interpretability, a causal learning paradigm is designed, allowing Obstet-LLM to generate clear and understandable explanations. Experimental results show that Obstet-LLM achieves an accuracy rate of over 90% in predicting SGA-LGA newborns, outperforming existing data-driven models. Moreover, by integrating domain-specific knowledge, Obstet-LLM provides actionable insights for clinicians, thereby enhancing the quality of prenatal care and reducing the risk of adverse pregnancy outcomes.