Translating findings from prenatal ultrasound reports into precise genetic risk assessments is a significant clinical challenge due to their complex, unstructured nature. Large language models (LLMs) offer a promising approach for automating this process.This study developed and evaluated a zero-shot innovative framework using DeepSeek-V3.2 to perform multi-dimensional classification of 254 abnormal ultrasound reports from a cohort of 4,256 pregnant women.This framework employs five distinct classification schemes—Standardized Terminology, Primary Classification, Anatomical System, Abnormality Count, and Severity. The LLM classifications were validated against expert annotations and correlated with amniocentesis-derived genetic outcomes for 251 cases. The code included the generation of Figures 2 and 3, evaluation of LLMs performance metrics, and statistical analyses in our research.