ClinLens: A Lightweight Coding Agent for Clinical Data Extraction from Medical Notes

A new preprint introduces ClinLens, a lightweight coding agent designed to extract structured clinical data from unstructured medical notes. The authors argue that existing NLP approaches for clinical information extraction are often heavy, task-specific, or require large labeled datasets. ClinLens takes a different path: it leverages large language models (LLMs) to synthesize small, interpretable programs—typically queries in a domain-specific language—that can be run against the text to extract targeted variables like medication dosages, lab values, or diagnoses.

The system works by first parsing a clinical note into a structured representation (e.g., sections and sentences), then using an LLM to generate a program that maps from that representation to the desired output. The generated program is executed by a lightweight runtime, making the extraction step transparent and debuggable. In evaluations on several standard clinical extraction benchmarks, ClinLens achieved competitive or superior F1 scores compared to fine-tuned models and larger black-box LLMs, while being far more resource-efficient.

What sets ClinLens apart is its emphasis on interpretability and low computational cost. Rather than relying on huge end-to-end neural models, it decomposes the task into a small, human-readable script. This is particularly valuable in healthcare settings where explainability and auditability are critical. The authors also show that the same approach can be adapted to other biomedical extraction tasks with minimal changes.

While the paper is still a preprint and likely needs validation on more diverse datasets, ClinLens represents a promising direction for building specialized, lightweight coding agents that solve real-world extraction problems without massive compute or data requirements.