Clinical documentation research · Kerala, India
Clinoria is a research initiative studying how multilingual clinical conversations become structured documentation that keeps its meaning: what was denied, what was resolved, who it was about, how certain it was, and which words support each line.
Synthetic data Prototype screenshot on a synthetic development case — no patient data. Photographs: Vision College (CC BY-SA 4.0) and Shixart1985 (CC BY 2.0), both via Wikimedia Commons. Modifications: cropped and resized.
We are not announcing that clinical AI is solved. We are publishing what we have built and what still fails.
The problem
Speech-to-text writes down sounds. Summarisation compresses detail into prose. Neither preserves what a clinician actually signs their name to: the exact denial, the exact timeline, the exact uncertainty. In medicine those distinctions change decisions.
Transcribe, then summarise
Extract facts, then render deterministically
Our approach
The transcript is treated as evidence, not as a draft. Clinoria's job is to decide what each part of the conversation means clinically, record that decision in a structure you can read, and only then write a note from it.
Consultation audio or focused dictation, captured with an explicit statement of which transcription route is used.
Speaker-attributed text. Treated as the evidence base for everything downstream, never edited silently.
Concepts with a closed status vocabulary, subject, laterality, temporality and a source span.
A note rendered from the fact graph under a department template, validated before it can be reviewed.
No cloud LLM drafts the note Drafting is a deterministic renderer over the validated graph. There is no step in which a language model writes free prose into a patient record.
Clinical intelligence
Each of these is a field in the fact graph, not a nuance left to a summariser's judgement. Where the extractor cannot populate a field reliably yet, we say so.
A denied finding is stored as ABSENT, which is a different fact from a finding that was never mentioned.
Historical and current states coexist: fever resolved last week and no fever today are both retained.
“My mother has diabetes” is attributed to the mother, not the patient.
Hedged statements are meant to stay hedged rather than hardening into assertions.
When a patient denies and later affirms the same symptom, both facts are kept side by side.
Every fact points at the transcript span it came from, down to the turn.
Where this is honest and where it is not
Negation, temporality, subject attribution, contradiction and provenance are
implemented and tested. Uncertainty is defined in the schema but is not
yet populated reliably by the extractor — hedged language currently
arrives as null. We describe the architecture here and the defects
on the research page, because a capability list without
a failure list is marketing.
Multilingual clinical research
Kerala clinicians do not speak one language at a time. A consultation moves between Malayalam, English and a Malayalam written in Latin characters. That mixing is the research problem, and it is not solved here.
Demonstrated
Known to fail
Planned research
The working prototype
These are real screens from the prototype, run on synthetic cases. We have not published a public demo: the extraction defects on the research page are reachable from a free-text box, and a live endpoint would show them without anyone present to explain them.
A demonstration takes about twenty minutes and runs both the text and the audio paths separately, because they behave differently.
Research and evaluation
Evaluation here is engineering evaluation: synthetic case sets, adversarial inputs, unit tests and a written defect register. It is not a clinical trial, and no independent validation exists yet.
Method
A 12-case ultrasound suite used to detect regressions. Internal developer evaluation set — not an efficacy study, and not a published benchmark.
Method
Negated red flags, mid-sentence corrections, compound family history and hedged language, written to break the extractor rather than flatter it.
Finding
Polarity inversion on code-switched negation, fabricated laterality passed by the validator, and dropped family-history clauses are all reproducible today.
For clinicians
The single most useful contribution a Kerala clinician can make is not praise. It is consented, de-identified consultation transcripts with your own gold-standard documentation for each — so that “correct” stops meaning “the engineer thought it looked right”.
What participation involves
Research collaboration
Demonstrations run live with the engineering team on synthetic data. Text and audio are demonstrated separately, because they behave differently.
Assistive draft only — every note requires credentialed clinician review and sign-off. Pre-clinical. Not a medical device. Please do not send patient data through this website.