Taking the hospital needlestick injury report as the application scenario,to construct a comprehensive assessment framework of "theme quality-interpretability-action transformation-technical characteristics" to compare the potential Dirichlet scores,the differences between the two topic modeling methods,LDA and the Deep Learning Topic Framework Based on Pre-trained Language Models(BERTopic),in risk identification and management decision support,and input the theme-driven intervention suggestions.

A total of 160 needlestick injury reports were collected from a tertiary hospital between January 2018 and September 2025.After preprocessing including Chinese word segmentation,stop-word removal,and synonym merging,LDA and BERTopic models were trained separately.Model comparison was conducted using topic coherence,expert blind evaluation,and technical indicators such as training time,parameter sensitivity,and degree of automation.The resulting topics were then mapped onto a tiered intervention framework to formulate actionable countermeasures.

Both models identified key risk scenarios including blood specimen collection, needle withdrawal/needle handling,and complex procedures in special departments.LDA generated five macro-level topics,suitable for rapid overview and routine monitoring.BERTopic generated four semantically coherent topics and one noise cluster, which more clearly delineated actionable scenarios such as blood collection chain risks among intern nurses,systemic failure in post-needle processing,and exposure in high-infection-risk cases.The topic coherence scores for BERTopic and LDA were 0.587 and 0.452, respectively.The overall expert rating scores were 4.555 and 3.944,respectively.Training times were 182 seconds and 45 seconds,respectively.

BERTopic demonstrates superior advantages in semantic depth and action translatability, making it suitable for root cause analysis and targeted improvement initiatives.LDA is better suited for rapid daily screening and trend monitoring.This study recommends a tiered application strategy of "routine LDA monitoring plus periodic BERTopic deep dives" in hospitals.The topic results can be transformed into structured interventions and evaluation indicators targeting populations,workflows,and devices,thereby forming a closed-loop management cycle of risk identification-intervention-review.