Context over time
We study how AI reads what is current, what is historical, and how health data changes over the years.
Research program
Research & Safety
How we study and improve the way AI works with health information.
Kanzen is a non-profit project of the KanzenKai Foundation, built to help people organise and better understand their own health information. Part of our work is to study, systematically, how consistently and clearly AI can work with health data that accumulates over years.
Our approach combines structured data, clear sources, checks on how the AI behaves, and comparisons between Bulgarian and English.
Our approach
AI that works with context, not just with a single question
Health information means something in context. A lab result, a medicine or a symptom can look different when you read it against earlier results, changes over time and other entries.
This is why Kanzen is aimed at longitudinal health data — information that accumulates and changes over time.
Clear sources
The goal is for answers to stay tied to the information that is actually there, rather than to assumptions.
Language people can use
We study how complex health information can be explained in a way that is clear and genuinely useful.
What we research
What we research
The KanzenKai research program focuses on how AI can work more reliably with personal health records, and how such systems can be evaluated in a measurable, reproducible way.
Grounding
Answers based on the information that is there
We study whether what the AI states can be traced back to specific entries and facts in the record.
Temporal reasoning
Reading change over time
We check whether the AI tells apart old and recent results, medicines that were started and stopped, and other changes.
Uncertainty
Being clear when the information is not enough
We study whether the system says plainly that there is not enough information for a confident answer.
Multilingual performance
Bulgarian and English
We compare how the AI behaves on equivalent health scenarios in Bulgarian and in English.
Consistency
Stable behaviour
We check whether similar questions over similar data lead to stable, predictable results.
Method
How the research works
For the initial studies we use synthetic health records, written specifically for testing. They imitate realistic situations without containing the data of real people.
We build scenarios with lab results, medicines, symptoms, measurements and events spread over time. Then we look at how the AI copes when the information is clear, incomplete, contradictory or changing.
- 1Synthetic record
- 2Question
- 3AI response
- 4Evaluation
- 5Improvement
This lets us compare different models, settings and approaches against the same set of controlled tests.
Evaluation
How we measure quality
Every question we research needs a measure that another team could apply to the same material and get the same number.
| Measure | What it means |
|---|---|
| Grounding accuracy | Whether the answer rests on the health entries that are actually present. |
| Source accuracy | Whether the entry that is cited really supports the statement. |
| Temporal accuracy | Whether current and historical information are told apart correctly. |
| Uncertainty handling | Whether the AI shows plainly that the information is not sufficient. |
| Consistency | Whether similar cases lead to similar behaviour. |
| Multilingual robustness | Whether the quality holds between Bulgarian and English. |
In the technical evaluations we also measure unsupported claim rate — how often a statement cannot be supported by the data that was provided.
Scope
Kanzen helps you understand your own information
Kanzen was built as an assistant for working with your own health information — to organise it, explain it and put it in context.
It is not meant to replace a doctor or any other medical professional. For important medical decisions, professional judgement remains what matters most.
This page describes what we research and plan to evaluate. It is not a description of capabilities the product already guarantees.
Research without real patient data
The initial research program uses synthetic records, written specifically for testing.
That lets us evaluate how the AI behaves without these experiments needing identifying health data of real people.
Open research
Research that can be useful beyond Kanzen
Our aim is not only to improve Kanzen. Where it is practical, we plan to publish methodology, synthetic benchmark components, aggregate results and observations that other teams working on responsible health-AI systems can use.
One particular focus is evaluation of AI in Bulgarian — an area with far less public data and far fewer benchmarks than English.
Current research program
Current Research Program
Project
Safety, Reliability and Grounding of AI over Longitudinal Personal Health Records
Research focus
- grounding to source health records
- temporal reasoning
- uncertainty handling
- Bulgarian compared with English
- consistency under different phrasings
- incomplete and conflicting information
- evaluation methodology for patient-facing AI systems
Planned duration
6 months
Status
Research program in preparation / evaluation phase
The program is run and funded by the KanzenKai Foundation. We do not claim funding or endorsement from any AI provider.
For researchers · technical details
A short note on how the evaluation is built, for readers who work on the same problems.
- Synthetic benchmark construction: records are generated from templates with controlled timelines, units and edge cases, so every case has a known expected answer.
- Repeated trials: each case is run several times, so a single lucky or unlucky response does not become a result.
- Perturbation testing: the same case is rephrased, reordered and partially removed, to see how stable the behaviour is.
- Unsupported-claim measurement: each statement in an answer is checked against the entries that were provided.
- Source attribution: we record whether a cited entry actually supports the statement it is attached to.
- Multilingual paired cases: every case exists in Bulgarian and English with the same content, so the two can be compared directly.
- Evaluation harness: cases, runs and scores are kept in a reproducible pipeline, so a result can be repeated later.
Collaborations
Research collaborations
KanzenKai is open to working with universities, medical professionals, health informatics researchers and organisations focused on safe and useful applications of AI in healthcare.
Kanzen is a non-profit personal health information project by the KanzenKai Foundation. The research described on this page focuses on AI evaluation and health-information assistance, and does not constitute clinical research, diagnosis, treatment or medical advice.