Adaptive Fast-Slow Large Language Model Framework for Multidimensional Classification of Prenatal Ultrasound Reports:...

Phenotype-driven prenatal diagnosis relies on the precise correlation between ultrasound findings and genetic outcomes; however, this process is hindered by the unstructured nature of clinical ultrasound reports. While large language models (LLMs) hold the potential to address this challenge, their specific application in this domain remains systematically underexplored.

To establish an effective LLM implementation framework for the clinical multidimensional classification of prenatal ultrasound reports, we evaluated the open-source DeepSeek-V3.2 family on real-world anomalous reports-covering both factual and subjective categories-while integrating retrieval-augmented generation (RAG) and chain-of-thought (CoT) reasoning.

From a cohort of 4256 pregnancies, we extracted 254 reports with fetal anomalies. We comprehensively evaluated both the high-speed base model (DeepSeek-V3.2-B) and the reasoning-enhanced model (DeepSeek-V3.2-R) across all 5 classification dimensions, comprising 4 factual extraction tasks-primary classification, standardized terminology, anatomical system, and abnormality count-and 1 subjective severity assessment. We further explicitly evaluated the efficacy of RAG for the subjective tasks. Finally, to validate the clinical utility of this approach, we performed a correlation analysis between the expert-validated multidimensional phenotypic profiles and definitive genetic outcomes derived from amniocentesis.

While V3.2-B achieved high efficiency in factual tasks (accuracy and F1-score >90%), it underperformed in subjective severity grading (56.6% accuracy), exhibiting a recall of 0 for minor anomalies. Crucially, while RAG significantly improved both models' performance on internal retrieval datasets (P<.05), this benefit did not generalize to external test datasets (P>.05). In contrast, the V3.2-R model utilizing CoT reasoning achieved superior robustness (86% accuracy and F1-score=0.75) on external data without RAG; notably, introducing RAG to V3.2-R degraded performance to 81%, suggesting potential noise interference. Clinical validation against amniocentesis outcomes confirmed that accurate multidimensional phenotypic profiles significantly stratified pathogenic genetic risks.

The rapid base models are efficient for factual classification, and RAG enhances performance on data similar to the knowledge base, whereas CoT is indispensable for subjective assessment. Within the constraints of our dataset and current retrieval implementation, CoT proved more robust than RAG for subjective assessment. However, this finding is specifically tied to our experimental setup and should not be generalized as a universal conclusion. We recommend clinically adopting this adaptive "fast-slow" LLM framework to efficiently perform the multidimensional classification of prenatal ultrasound anomalies. This privacy-preserving, locally deployable solution provides a scalable path to accelerate phenotype-genotype research and optimize invasive diagnostic decision-making.