A Tragedy in One Word (and Two Languages)

The gap nobody checks for

Every hospital has language access. Interpreters show up for appointments. Phone and video services cover 240 languages. The patient journey for limited-English-proficient patients looks parallel to everyone else’s — right up until they leave.

At discharge, an interpreter reads the instructions aloud. The patient nods. Then they go home with an English-only document. In a crisis, patients retain about 20% of what they’re told verbally. The written instructions are supposed to be the fallback.

For 25 million Americans, that fallback is in the wrong language.

The cost of one word

Steve Labkoff, our co-host, shared a story from his residency at Cooper Hospital. A Hispanic mother brought her baby in with lice and was sent home with lindane shampoo. The instructions said to apply it “once” a day. In Spanish, “once” means eleven. She applied it eleven times daily. The baby died from neurotoxic exposure.

Giovanni brought his own case — the Tran family, where a Vietnamese child died after the older brother interpreted and missed a critical warning about a contraindicated medication on the discharge paperwork. UC Berkeley and the National Health Law Program have documented more than 35 similar cases.

Three “solutions” that don’t work

When we asked how hospitals currently handle this, the answers were consistent: an interpreter reads the document aloud (patient leaves with the English copy), an in-house team translates into Spanish (costly, one language, takes over an hour), or a vendor returns the translated document in 24 to 48 hours — long after the patient needed it.

The surprise: Epic’s language module translates field labels like “Prescription:” and “Therapy:” into other languages. The actual clinical content inside those fields stays in English.

How FETCH works

Donatelli’s solution runs inside Epic. A clinician selects the patient’s language, hits print, and gets a bilingual document in 15 minutes. The process has three layers: a translation memory handles 75-85% of familiar medical text, a client-specific AI (closed, US-hosted, not a consumer tool) fills the gaps, and a human reviewer finalizes and retrains both systems.

The patent covers a PHI redaction method that strips protected health information before translation, keeps PHI on US soil, and reassembles the document afterward. This lets Donatelli use a global pool of linguists without violating HIPAA.

Translators become trainers

We asked what happens if regulations drop the human-in-the-loop requirement. Donatelli’s answer was immediate: “There’s a future for my translators. They’re no longer translating documents — now training AI to translate documents.”

He built FETCH so the human layer can scale up or down. Remove oversight? Turnaround drops to milliseconds. Tighten it? The architecture already supports that. And his company accepts malpractice liability for translation errors — a position most AI vendors won’t take.

The CIO blind spot

Donatelli’s challenge to healthcare executives is blunt: “You think you have this solved. You don’t. It doesn’t exist.”

Ask your team whether discharge instructions get translated. They’ll say yes. Ask how. That’s where the gap lives.

Listen to the full conversation with Giovanni Donatelli on Practical AI in Healthcare.