Medical calculators convert raw patient observations into clinically actionable values and scores. Applying them directly to free‑form patient notes requires precise variable extraction and exact execution of complex formulas, scoring rules, and output formats. Pure LLM prompting is unreliable because rule recall, parameter extraction, arithmetic reasoning, and formatting are tangled in a single generative step. To overcome this, we introduce a tool‑augmented LLM framework. The framework first retrieves and structures calculator specifications, then generates reusable executable tools, and finally uses a constrained extractor to map each patient note to inputs ready for the calculator. The tools perform deterministic computation, separating clinical language understanding from rule execution and arithmetic. Evaluated on the MedCalc‑Bench dataset, the approach achieves 86.73% final‑answer accuracy, surpassing one‑shot prompting and supervised fine‑tuning baselines by 24.0 and 17.27 percentage points respectively. Ablation studies show that gains stem mainly from rule‑grounded tool synthesis and schema‑constrained extraction, while remaining failures concentrate on clinical‑semantic extraction and benchmark‑specific output contracts. By grounding calculator logic in executable tools, the method reduces arithmetic hallucinations and improves the computational reliability of clinical outputs.
Review