🗣 Back-Translation Verification for Medical AI Simulator

Forward Translation of Clinical Text

Medical AI translates source text into the target language.

• English: Source language (clinical documentation)
• Spanish: Target language (patient-facing translation)
• 6: Phrases translated (dosage and safety lines)
• Adjustable: Model quality (basic to advanced slider)

Why translate medical text at all

Patients need instructions in their own language.

Where the AI model fits in

A neural model converts source sentences automatically.

A single mistranslated dosage can cause real patient harm.

What can go wrong early

Rare terms and dosages are easy to mistranslate.

Reverse Translation to Source Language

The translated output is fed back through translation again.

• ES→EN: Reverse direction (back to source language)
• 2: Round-trip hops (forward then reverse pass)
• Yes: Independent pass (no memory of original text)
• Cumulative: Drift risk (errors compound each hop)

The back-translation technique

Translated text is translated back without seeing the original.

Why independence matters

A fresh translation exposes errors the first pass hid.

Round-trip translation is a cheap proxy for human review.

Limits of the method

Back-translation catches meaning drift, not stylistic changes.

Sentence-Level Comparison Analysis

Back-translated sentences are aligned against the original wording.

• Semantic: Similarity metric (meaning, not just wording)
• Phrase: Alignment unit (sentence-level comparison)
• Embedding: Comparison method (vector distance scoring)
• 80%: Pass threshold (minimum similarity score)

Aligning sentence pairs

Each phrase is matched to its back-translated counterpart.

Scoring semantic similarity

Embeddings measure how close two meanings really are.

Wording can differ while meaning stays fully intact.

Setting a pass threshold

Below the threshold, a phrase is marked suspect.

Flagging Meaning-Altering Discrepancies

Phrase-level mismatches reveal where meaning may have shifted.

• 4: Discrepancy types (dose, timing, severity, negation)
• Critical: Danger class (dosage and negation errors)
• Red: Flag color (highlighted mismatched phrases)
• <1s: Detection speed (per phrase pair)

Types of discrepancy

Dosage, timing, severity, and negation errors matter most.

Why negation is dangerous

Dropping a single "not" can reverse an instruction.

Negation and numeral errors cause the most harm.

Surfacing flags for review

Mismatched phrases are highlighted directly for translators.

Approval or Human Review Verdict

A pass or review verdict closes the safety loop.

• 2: Verdict states (approved or review needed)
• Required: Human reviewer (when similarity is low)
• Logged: Audit trail (every verdict recorded)
• Hard stop: Deployment gate (blocks unsafe translations)

Reaching a verdict

High similarity across phrases yields automatic approval.

Routing to human review

Low similarity sends the translation to a human.

Human review remains the final safety net.

Closing the safety loop

Every verdict is logged for audit and retraining.