Abstract & aims
Three drafts shaped for where they go: a conference abstract, an IRB pre-submission, and a grant specific-aims page. The counts are read live from the evidence bank, so the text tracks the shipped app. It claims no outcomes, because we have measured none yet.
A citation-locked, deterministic clinical decision-support tool for food-as-medicine that keeps generative AI out of the medical path
Structured abstract, roughly 300 words. Shape to the venue's word limit.
Background
Dietary guidance for chronic disease is siloed across condition-specific guidelines, leaving clinicians to reconcile overlapping and sometimes conflicting advice for multimorbid patients. Tools that use large language models to fill this gap are documented to deviate from guidelines in meal planning, and their output cannot be traced to a source. We describe a decision-support tool designed to remove both failure modes.
Methods
The system is a deterministic rules engine over a versioned, citation-locked evidence bank (24 rules drawn from 27 professional-society, federal, and USDA sources, each anchored by a verbatim supporting quote; all 24 rules carry a dated, named clinician sign-off). Population safety gates withhold automated output and route to a named clinician where guideline diet advice does not apply (for example pregnancy, advanced kidney disease, or eating-disorder risk). Drug and food interaction rules adjust guidance to the patient's medication list across 17 medications. A 10-year cardiovascular risk estimate is computed in-browser from the published American Heart Association PREVENT equations. Every nutrition value carries USDA FoodData Central provenance. Identical input produces identical output, verifiable by SHA-256 hash. No language model sits in the medical path.
Results
10 conditions are fully modeled with 24 cited, clinician-signed-off rules from 27 sources; 23 recipes carry per-nutrient provenance to USDA records. 9 demonstration patients exercise tailoring across conditions and medications, conflict resolution between competing nutrient targets, and refusal behavior for populations the engine declines to plan for. A built-in measurement loop records diet quality (the validated MEDAS screener), self-reported adherence, and cardiometabolic markers over time. No clinical outcomes are reported; this is a system description.
Conclusion
Traceability to a named source with quoted text aligns with the FD&C Act section 520(o)(1)(E) criterion that a clinician be able to independently review the basis for a recommendation, a claim a black-box model cannot make. The deterministic design also makes the intervention exactly reproducible for evaluation, since a stored profile and bank version regenerate the identical plan. A supervised pilot to measure feasibility and cardiometabolic change is the next step.
Counts are read live from evidence bank version 0.3.0. Trim to the venue word limit and confirm every figure against the shipped app before submitting.
Draft language only. Verify every sentence, figure, and claim against the shipped app and the target venue's template before submitting.