Methodology

Language understanding never does arithmetic

Large language models are unreliable at clinical calculation: in the MedCalc-Bench benchmark (NeurIPS 2024), frontier models made frequent formula and arithmetic errors. AskCalc therefore uses language technology only to choose a calculator and read the values out of your text. A rule-based parser handles most questions instantly; an optional small language model (Qwen3, 1.7B parameters) runs inside your browser through WebGPU for freer phrasing. The calculation itself is plain, deterministic code.

Safety rules

• The model may report that a finding is present, never that it is absent: items you did not mention are asked, not assumed.
• Numbers without units are checked against plausible ranges; a creatinine of 132 is read as µmol/L, and implausible values are refused.
• Every result lists the inputs it used; values filled by the model are flagged for you to verify.

Testing

Each calculator ships with test cases whose expected values were derived independently from the original publication (worked examples or hand calculation). The full suite runs on every build; a single failure blocks publication.

Sources

Formulas and point values follow the original derivation paper or the current guideline, linked on every page. Where sources disagree, the page says which version is used.