Hallucination Guardrails

Forces all generated medication names to resolve against the RxNorm normalized naming system. Any LLM output that does not map to a valid RxNorm Concept Unique Identifier (RxCUI) is automatically rejected.

• Validates against the full RxNorm graph including brand names, generics, and ingredient-level concepts
• Prevents generation of near-miss drug names like 'metformine' instead of 'metformin'
• Links to Structured Product Labeling (SPL) for machine-readable drug monograph verification

Applies hard constraints on generated dosage values using FDA Structured Product Labeling maximum daily dose limits and Renal Dose Adjustment thresholds. Any output exceeding established therapeutic ranges triggers automatic suppression.

• Validates strength, frequency, and route against drug-specific monographs
• Integrates patient-specific estimated glomerular filtration rate (eGFR) for renal dosing guardrails
• Prevents generation of lethal decimal-place errors like '10.0 mg' instead of '1.0 mg'

Employs Active Ingredient Matching algorithms to detect when an LLM generates a medication that is therapeutically equivalent to an existing order. Resolves brand-generic pairs to their base RxNorm ingredient to prevent Duplicate Therapy Alerts.

• Cross-references all generated entries against the patient's current medication list
• Flags structural similarities using Allergen Cross-Reactivity logic
• Prevents hallucinated duplicate orders that could lead to overdose

Applies Temporal Reasoning rules to verify that generated medication timelines are logically coherent. Rejects outputs where start dates precede discontinuation dates or where a medication is ordered after a documented severe adverse reaction.

• Sequences clinical events chronologically using timestamp metadata
• Validates against the Medication History Longitudinal Record for cross-encounter consistency
• Prevents impossible timelines like a medication 'started' after patient discharge

Implements Confidence Thresholding gates that route LLM outputs to Human-in-the-Loop (HITL) review when the model's prediction score falls below a predefined certainty level. High-confidence extractions proceed automatically; low-confidence outputs are quarantined.

• Uses calibrated probability scores rather than raw logits
• Integrates with clinical review interfaces for pharmacist verification
• Optimizes the balance between automation throughput and patient safety

Integrates NegEx Algorithm patterns and contextual negation detection to prevent the LLM from extracting medications that are explicitly documented as 'not taken,' 'discontinued,' or 'allergic to.'

• Extends beyond regex to transformer-based negation detection using ClinicalBERT embeddings
• Distinguishes between 'patient denies taking' and 'patient reports taking'
• Prevents hallucinated active medications from negated clinical statements

An architectural pattern that grounds a language model's output by first retrieving factual medication data from a vector store or knowledge base before generation. In medication reconciliation, RAG queries RxNorm or a hospital formulary to provide the model with verified drug names, dosages, and routes, significantly reducing the probability of generating plausible-sounding but non-existent medications. The retriever acts as a factual anchor, constraining the generator's output space to only attested clinical entities.

A probabilistic gate that routes AI-extracted medication data for human review only when the model's prediction score falls below a predefined certainty level. For example, if a model extracts 'atorvastatin 40mg' with 0.97 confidence, it passes automatically; if it extracts 'atorvastatin 400mg' with 0.62 confidence, it is flagged for pharmacist verification. This optimizes the balance between automation throughput and patient safety by ensuring low-certainty outputs—where hallucinations are most likely—never reach the patient record unchecked.

A deterministic post-processing step that maps every extracted medication string to a unique RxNorm Concept Unique Identifier (RxCUI). If a model generates 'Lipitor' or a misspelled variant, the normalization layer resolves it to RxCUI 153165 (atorvastatin). Any generated string that fails to match a valid RxNorm concept is rejected outright, creating a hard guardrail: no medication name enters the system unless it exists in the authoritative national drug vocabulary.

The algorithmic technique of linking brand-name and generic drug products by resolving their chemical constituents to a common base compound. This prevents duplicate therapy errors where a model might hallucinate a new brand variant of an existing generic. The guardrail cross-references the active moiety—e.g., mapping both 'Tylenol' and a hallucinated 'Tylenol XR' to 'acetaminophen'—and triggers a duplicate alert if the patient already has an active order for the same ingredient.

The computational process of converting disparate representations of medication strength and frequency into a standardized, comparable format. A hallucinated dosage like 'metformin 1000mg twice daily' is normalized to 'metformin 2000mg/day' and validated against the drug's FDA Structured Product Labeling (SPL) maximum daily dose. Any normalized value exceeding the SPL-defined safety ceiling is automatically rejected, preventing clinically dangerous dosage hallucinations from reaching the prescriber.

The mechanism of explicitly linking each extracted medication entry back to the specific sentence, document, or database field from which it was derived. When a model generates a medication claim, the guardrail system requires a citation pointer to the source text. If no source span can be aligned with the generated output—indicating the model fabricated the medication from its parametric knowledge rather than the patient's record—the entry is suppressed. This enforces strict evidentiary grounding for every clinical assertion.