AI Suggest — medication safety that can say no.

When a doctor clicks 'AI Suggest', the system performs clinical reasoning on the patient's transcript, current medications, conditions, allergies, and organ function data. AI models identify candidate medications from a comprehensive formulary. Each candidate is then passed through 24 deterministic safety gates — rule-based checks that cannot hallucinate — covering DDI, drug-disease contraindications, renal/hepatic dosing, QTc prolongation, allergy cross-reactivity, and more. Surviving candidates are ranked by clinical evidence tier and returned to the doctor as a reviewed, labelled list.

A deterministic safety gate is a rule-based check that produces a guaranteed outcome (PASS, CAUTION, or CRITICAL) based on verified clinical data — it does not use probability or estimation. This matters because AI language models can confidently suggest dangerous drug combinations in edge cases. ELLEXMED's gates intercept those suggestions before they reach the clinician, regardless of what the underlying AI model outputs.

MedQA is a multiple-choice benchmark derived from US Medical Licensing Examination questions, and frontier models already score around 90% on it. A high score would say something about the underlying model, not about whether a prescribing system is safe. What matters for prescribing is not how often the model is right, but what happens when it is wrong — so we invest in the deterministic layer that rejects unsafe candidates regardless of model accuracy, and we test that layer on refusal rather than on trivia.

Every medication candidate is screened by 24 deterministic gates that do not consult the language model, backed by an automated suite covering gate semantics that runs on every change. Alongside accuracy, we assert monotonic safety properties that must hold regardless of which model is behind them — raising patient acuity never loosens the mode, and crossing age 65 never removes a Beers-criteria flag. We are also building adversarial evaluation on the UK AI Safety Institute's open-source Inspect framework, to measure refusal under pressure rather than average accuracy. That work is in progress and we do not claim results from it yet.

ELLEXMED checks pairwise pharmacokinetic interactions from curated pharmacological databases, plus CYP450 metabolic pathway conflicts (which cause most clinically significant DDIs). A neural vector-similarity layer provides additional coverage for interactions not present in pairwise databases, using 1,084 drug-pair vectors calibrated at ALERT threshold ≥0.70 and CAUTION threshold ≥0.67.

Every CRITICAL or CAUTION override is acknowledged by the clinician and timestamped in ELLEXMED's immutable audit log. The override records the gate that was triggered, the drug involved, and the clinician's identity. This creates a verifiable clinical decision trail for compliance and review purposes.

No. Every AI Suggest output is explicitly a draft for clinician review — the system informs, the doctor decides. The AI ranks and validates candidates; the clinician has full authority to accept, modify, or reject every suggestion. ELLEXMED is designed as a clinical copilot, not an autonomous prescribing system.

AI Suggest checks renal function (eGFR-based dose adjustment per KDIGO guidelines), hepatic safety (Child-Pugh classification for hepatic metabolism), and QTc prolongation risk from FDA drug label data. Medications requiring dose reduction for renal or hepatic impairment are automatically flagged or adjusted.