Impact of FHIR Data Serialization Strategy on LLM Medication Reconciliation
[论文评述] Serialisation Strategy Matters: How FHIR Data Format Affects LLM Medication ...
Summary
This study compared the impact of different serialization strategies when inputting FHIR-compliant patient records into Large Language Models (LLMs) for medication reconciliation during healthcare transitions. The results showed that the optimal input format varies depending on the model and data volume, emphasizing the critical importance of maintaining high recall.
Details
This research systematically evaluated the performance of LLMs on Medication Reconciliation using a cohort of 200 synthetic patients with FHIR R4 records. Given that medication reconciliation is a high-risk process in healthcare transitions, LLM assistance is highly anticipated. The study generated text inputs using four distinct serialization strategies from the FHIR data: ①Raw JSON (baseline), ②Markdown Table, ③Clinical Narrative, and ④Chronological Timeline. These were tested across five open-weight models (including Phi-3.5-mini). Notably, the 'Clinical Narrative' strategy showed performance improvements by explicitly separating active medications from historical ones, making it the most recommended format for 8B class models. Conversely, large models (70B) performed slightly better with Raw JSON. Furthermore, the failure mode across all strategies was found to be 'Omission' rather than 'Hallucination,' suggesting that system completeness is paramount for clinical safety audits. The research also demonstrated that domain pre-training alone is insufficient for structured extraction tasks; strong instruction following capability is essential. In conclusion, selecting the optimal data format (serialization strategy) and prioritizing recall are crucial considerations for deploying LLMs in clinical settings.
Original content copyright by respective publishers