Methods, evidence and limits

How SentinelDose reaches a finding, what it rests on, and what it does not do. Rule set version 1.0.0.

Two layers

Rule engine. A deterministic engine written in TypeScript runs in your browser. It recognizes 545 medicines and 97 foods, supplements, substances and situations. Each carries curated properties: metabolic pathways, enzyme inhibition and induction strength, QT liability, anticholinergic burden score, and graded effects on bleeding, potassium, blood pressure, sedation and more. From these it computes cumulative burdens and pharmacokinetic interactions, and it applies 77 individually written interaction rules and 17 prescribing-cascade patterns. The same input always gives the same output.

AI synthesis (Pro, optional). A language model receives the de-identified case with the rule findings. It re-ranks them for the individual patient, adds short notes, comments on entries the rules did not recognize, and may add concerns. Anything the model adds is labeled “AI-added, verify” and carries lookup links only, because the model is not permitted to supply its own citations.

Drugs outside the curated list

When a name is not in the curated list, the app can look it up in RxNorm, the drug-name database of the US National Library of Medicine. A brand is resolved to its ingredients. An ingredient that is not curated is placed in its WHO ATC drug class and receives that class's rules only, and every finding that depends on it is graded no higher than “probable” and says so.

The FDA label cross-check then fetches the current FDA-approved label for each medicine from openFDA and searches its interaction, contraindication and boxed-warning sections for the other items on the list. Matches are shown as quotations with a link to the full label. They are not rated, and a mention can also state that no interaction exists.

Evidence strength on every finding

  • Established. Supported by regulatory labeling, a clinical guideline or controlled studies.
  • Probable. Consistent pharmacology with observational data, case series or expert consensus.
  • Theoretical. Based on mechanism. Clinical evidence is limited or conflicting, and the finding says so.

Each finding links to its sources. Links marked “Look up” are searches of FDA labeling (DailyMed) and PubMed for the agents involved, offered so you can check the current wording yourself. Every reference in the list below was checked to resolve to the named document when this version was built.

Privacy

Rule-based reviews are computed in the browser. The optional NIH and FDA lookups send drug names, and nothing else, straight from your browser to those agencies; they can be switched off under the Analyze button. Our server receives only counts and rule identifiers, so that the accuracy dashboard can show which rules fire and how they are rated. Feedback sends the rule identifier and the drug names in that finding. AI synthesis sends the de-identified case (age, sex, weight, labs, conditions and the lists you typed) to the model provider and stores nothing. Do not enter names, dates of birth or record numbers. Use with identifiable patient data requires your organization's own privacy and security review.

Limits

  • The rule set is curated, not exhaustive. Absence of a finding is not evidence of safety.
  • It has not been clinically validated in a prospective study and is not a cleared medical device.
  • Coverage is thin for oncology, antiretroviral, pediatric and specialty medicines. Pregnancy and lactation are not assessed.
  • It does not check indication, dose adequacy, allergies, pharmacogenomics or missing therapy.
  • Renal thresholds paraphrase labeling and prompt a dose check; they do not replace it.
  • Perioperative guidance follows the cited guidelines; local protocols and the surgical and anesthesia teams take precedence.
  • Guidelines change. Each rule cites the version it was written against.

References (55)

Guidelines

Regulators

Trials and cohort studies

Reviews

Reference databases