Evidence · medicine
Differential diagnosis
with a reason for every ruling
Medicine is one of the domains we built the engine against, because a wrong call costs something here and the right call has to be defensible. A patient describes symptoms. FROS checks 30,981 diagnoses against machine-checked rules derived from medical knowledge bases, and returns a ranked differential with exclusion certificates showing why each diagnosis was ruled in or out. Every ruling is inspectable, and it is one piece of a larger system. See how the engine keeps a model coherent over a long task.
See it in action
FROS DDx Engine
A patient describes crushing chest pain, sweating, and nausea. FROS identifies the clinical findings, weighs all candidate diagnoses, and returns a ranked differential. It runs from the patient's description alone.
By the numbers
Verified medical knowledge
Each with machine-checked rules derived from medical knowledge bases.
78.4% on out-of-sample USMLE / MedQA questions, engine-only. 100% on the curated vignette set used to tune the rules, the easier, in-sample number.
The part that checks a rule against a patient's findings is machine-checked to compute correctly given its inputs. That covers the logic; the clinician still reads and judges every certificate.
How it differs
Every ruling out comes with a certificate
AI-based diagnosis
Neural models produce ranked lists with probability scores. "Myocardial infarction: 73% likely." The score arrives bare: the reasoning stays hidden, the alternatives it weighed go unrecorded, and tomorrow the same input can return a different answer.
FROS diagnosis
The engine tests each diagnosis against the patient's findings using machine-checked rules drawn from medical knowledge bases. Diagnoses are ranked by how much evidence supports them. Each exclusion comes with a certificate showing exactly which findings ruled it out. The same inputs always yield the same answer, with a full trace you can inspect.
Disambiguation in medicine
Words matter in clinical text
Medical language is full of ambiguity. "Cold" can mean temperature or illness; "discharge" can mean release from hospital or bodily fluid; "positive" can mean good news or a concerning test result.
FROS resolves these ambiguities before clinical reasoning begins. The same engine that achieves 94.5% on standard NLP benchmarks processes clinical text, so downstream diagnosis operates on resolved meanings. Every sense assignment and every diagnosis exclusion is produced by the deterministic engine, which owns the decision path end to end. That closes a specific failure mode: an ambiguous term steering the diagnosis toward the wrong match.