un ataque
PhrasebankpolysemyTier 3 · unreviewedCould mean
- seizure (neurological)
- panic attack (psychological-somatic)
Clinical meaning-loss detection
The words arrive. The meaning should, too.
Sandhiva detects possible meaning loss in multilingual clinical encounters and gives clinicians the next question to ask.
Encounter intelligenceES → EN
ListeningSimulated
Madrees
Anoche mi hijo tuvo un ataque.
The next question
“Did the body shake or stiffen?”
Illustrative prototype · unreviewed example · not for clinical use
01 / See the signal
One phrase. More than one meaning.
Follow a simulated conversation as Sandhiva connects a patient’s words to possible meanings and a clarifying question. Scroll to move through the encounter.
Illustrative research prototype. These unreviewed example cards are not for clinical use.
Sandhiva
Simulated encounters · not for clinical use
Listening
Cards appear as phrases are spoken
Clinicianen
What brings you in today?
Resolved on match
Madrees
Anoche mi hijo tuvo un ataque.
un ataque
PhrasebankpolysemyTier 3 · unreviewedCould mean
- seizure (neurological)
- panic attack (psychological-somatic)
Ask
- Did the body shake or stiffen?
- Was there loss of consciousness?
Evidence and documentation
Basis: the illustrative phrasebank entry for un ataque. This Tier 3 example is unreviewed; the listed meanings are possibilities, not findings about a patient.
Documentation template: Clarified “un ataque” with the patient and interpreter. Patient’s explanation: [record their response after review].
Illustration only. No record is created or changed.
These clinical actions are unavailable in the simulated replay.Interpreteren
He has a lot of phlegm in his chest since he got a bad flu.
flema
PhrasebankpolysemyTier 3 · unreviewedCould mean · respiratory
- productive cough / sputum
- chest congestion
Ask
- Is there a cough that brings something up, or a whistling sound when breathing?
Evidence and documentation
Basis: the illustrative phrasebank entry for flema. This Tier 3 example is unreviewed; the listed meanings are possibilities, not findings about a patient.
Documentation template: Clarified “flema” with the patient and interpreter. Patient’s explanation: [record their response after review].
Illustration only. No record is created or changed.
These clinical actions are unavailable in the simulated replay.
The words were translated. This is the part that was not.
02 / What is at stake
Every word carries clinical weight.
- 67.8M
- U.S. residents aged 5+ spoke a language other than English at home.U.S. Census Bureau · 2019 ACS
- 25.5M
- U.S. residents aged 5+ reported speaking English less than “very well.”MPI analysis · 2019 ACS
The cost of one misunderstood word
A case documented by AHRQ turned on the word
intoxicado
Misread as intoxicatedIntended here: feeling sick to the stomach.
The misinterpretation was associated with potentially preventable quadriplegia. Clarifying the meaning matters.
AHRQ PSNet · Language Barrier, 200603 / The role we play
A second look at meaning.
A clearer next step.
Notice the ambiguity.
Find expressions whose clinical meaning may be lost between the patient, the interpreter, and the record.
Ask a better question.
See possible meanings and focused questions that help the clinician clarify what the patient meant.
Keep the clinician in control.
Review the basis for each card and choose what belongs in the note after clarification.
Two surfaces. One engine.
Planned deployment path
At the bedside
Designed for a hospital tablet alongside phone or video interpretation, so a question can be asked in the moment.
Inside telehealth
Planned integration with existing caption and transcript streams, bringing the same cards into a virtual consultation.
04 / The architecture of understanding
Retrieval first.
Generation last.
Start with curated clinical language. Escalate unresolved, unfamiliar phrases to a reasoning model only when needed. Show the basis for every suggestion.
- 01
Clinical phrasebank
Patient expressions, possible meanings, and clarifying questions, organized for clinical review.
- 02
Contextual retrieval
Match the conversation to relevant entries using meaning and surrounding context.
- 03
Ambiguity review
Check whether a phrase could carry a clinically important meaning that the translation missed.
- 04
A question to ask
Present an interpretation card with possible meanings, questions, and its supporting basis.
Clinical knowledge compounds.
We’re starting with a focused Spanish–English phrasebank and a clinician-adjudicated evaluation set. Each validated expression adds reusable knowledge for future encounters.
A tier identifies how an entry was reviewed. It is not a diagnosis or a confidence score.
- Tier 1 · clinician-validated
Clinician-validated entries. The foundation we are building.
- Tier 2 · reviewed
Model-assisted entries, reviewed before activation.
- Tier 3 · unreviewed
Generated candidates awaiting review. Shown here only as illustrative examples.
05 / The horizon
Translation carries the words. Understanding carries the meaning.
Our long-term ambition is clinical interpretation that can measure and disclose its own failures. Today, we are building the meaning-loss detection and clinical evaluation that vision requires.
The path ahead: support human interpretation, verify machine interpretation, and ultimately pursue autonomous interpretation with clinical evidence and regulatory review at each step.
06 / Built with care
Clinical judgment stays human.
Sandhiva supports clarification. It does not replace an interpreter, act as a scribe, diagnose, assign disease probabilities, recommend treatment, or triage. Every card shows its basis for independent clinician review.
- Now · demonstration and evaluation
- Building with simulated encounters and a clinician-adjudicated evaluation set. No patient data is used in this demonstration.
- Next · clinical design partnerships
- A health-system pilot and a telehealth design partner are next milestones. Clinical pilots would require appropriate agreements, privacy controls, and review.
Start a conversation
Build the next layer of understanding.
We’re looking for health systems, telehealth teams, and interpreter partners to help shape what comes next.
Let’s talk