Health
Care and
Cultivation Back to the homepage

IHAI research programs

IHAI studies how technology, traditional knowledge, human relationships and community participation can work together to support whole-person Health Care and Cultivation (HCC). Eight research programs carry that work. Each card shows what we have built or studied; open it to see the methods, findings and limits.

Exploratory research and prototype testing. No clinical outcomes have been established. The 2027 pilot is an implementation study, not a clinical trial.

How the programs fit together

  1. CareStories learns what people say about their health and lives.
  2. WisdomMap organizes health knowledge so every statement can be traced.
  3. CareSignals tests whether AI notices when context changes meaning.
  4. MemoryBridge carries context safely from one conversation to the next.
  5. LifeCompass brings these into a person's first orientation to HCC.
  6. HumanLink and TimeTrust ask how the people who give care are strengthened and recognized.
  7. RestWell applies the same evidence standards to one health concern: sleep.

How to read the evidence labels

Descriptive
A count or structure produced from a real body of material. It describes what was found, not whether anything works.
Internal test (synthetic)
Software testing by our own team, using fictional cases and simulated inputs. It shows that a function behaves as designed, not that it helps people.
Exploratory
A small early study with people or constructed cases. It can suggest questions; it cannot confirm an effect.
Conceptual
A framework, design or set of standards. It explains an idea; it is not an empirical result.
Planned
Work that is proposed and has not started.

Whole-person care technologies

Featured program

LifeCompass

Orientation Package: HCC conversational intake, with an optional traditional life-course study

Helps a person tell their story in their own words and begin a personal orientation to HCC, with people deciding what happens next.

  • Internal test (synthetic)
  • Exploratory
  • 40 of 40 functional checks reported passing, synthetic cases
  • 10 participants in a preliminary usability survey
Explore LifeCompass
Who
People managing ongoing health concerns, their families, and the stewards and practitioners who support them.
What
A first conversation that captures the whole person, not only symptoms.
Why
Care that starts from a person's own account is less likely to miss what matters to them.
Where and when
At the start of HCC participation, in English or Traditional Chinese.
How
A conversational intake that keeps the person's own words, flags gaps and safety concerns, and prepares a record for human review.

How LifeCompass works

  1. Conversational intakePrototype
  2. Optional AI-suggested clarifying questionsPrototype
  3. Practitioner review and sign-offPrototype
  4. The person approves what is sharedPrototype
  5. Optional traditional life-course explorationResearch stage
  6. Whole-person reportPrototype tested; not yet deployed
  7. Ongoing human stewardshipPlanned

Prototype means built and tested with fictional cases; it is not in clinical use. The standard pathway does not ask for birth information. The traditional exploration is optional, needs separate consent and can be withdrawn. Two report formats have been built and tested with fictional cases but are not yet deployed: a standard whole-person report that needs no birth information, and the same report with traditional life-course insights, only at the person's request.

Study 1: Conversational intake and usability

What we did
Built a browser-based intake prototype in English and Traditional Chinese. It keeps a person's own words separate from interpretation, flags missing details and possible safety concerns, and prepares records for practitioner sign-off. A guardrail blocks diagnostic or reassuring wording in the optional AI questions. We ran a 40-check functional suite and a preliminary usability survey with 10 participants.
What we learned
All 40 checks are reported passing on the current version. The usability survey is preliminary: its results are under review and will be published with their method.
Limits
The functional checks used fictional cases and simulated model responses; live-model performance has not been verified. The usability survey is small and uncontrolled. Neither shows clinical effectiveness.
Next
Archive the versioned test logs for audit; publish the usability method and results after review.

Study 2: Traditional life-course exploration

What we did
Explored whether traditional life-course frameworks can prompt questions and themes a person finds meaningful, with every statement traceable to its source and checked by the person. Three exploratory cases; the supporting research software is under acceptance review.
What we learned
In these three cases, interpretations about present health did not match what participants reported about themselves. Three cases cannot show how accurate or useful traditional interpretations are in general. LifeCompass does not use them to assess health; whether they help people reflect on life history and personal strengths is an open question.
Limits
Three cases. A participant agreeing with a statement is not evidence of predictive accuracy. No personal details are published.
Next
A comparison against mismatched profiles is the first planned test. Fit between two people and change over time are possible future research directions. They are not validated and are not used for any decision, including steward assignment.

Open the HCC intake prototype (v0.2)

The demo shows the conversational intake and the steward and practitioner views. It is not the complete LifeCompass application: the whole-person report and the optional traditional exploration are tested separately and not yet deployed.

MemoryBridge

IHAI Memory Bridge v1.2.5, longitudinal context tool

Explores how important personal context can be carried safely from one conversation to the next.

  • Internal test (synthetic)
  • 49 of 49 synthetic test cases reported passing
Explore MemoryBridge
Who
Participants, and the stewards who follow them over time.
What
Continuity of context across encounters.
Why
People should not have to repeat their story at every visit.
Where and when
Between and across conversations, as circumstances change.
How
Each item is kept with its source and time; contradictions are surfaced, not overwritten; access is limited by the person's choices.
What we did
Synthetic test cases covering provenance, time-based controls, contradiction handling and access governance.
What we learned
49 of 49 cases are reported passing in internal testing. No real participant data were used.
Limits
Synthetic testing only. Not approved for clinical use, and work with real patient data is on hold pending governance review.
Next
Archive the full test report; connect to the intake only after review.

Evidence and knowledge research

CareStories

FYTCM 51-recording narrative representation study

Studies what people say about their health and lives, and what records keep of it.

  • Descriptive
  • 51 clinic recordings analyzed
  • 657 eligible source statements
Explore CareStories
Who
Patients at a holistic medicine clinic emphasizing individualized care, preventive health and long-term health cultivation, and their practitioners.
What
Whole-person information in people's own accounts: routines, emotions, caregiving and change over time.
Why
Much of what shapes health is said aloud but never written down.
Where and when
In clinic conversations recorded over time.
How
AI-assisted extraction of source-linked statements, then a comparison of how different record formats hold them.
What we did
An exploratory analysis inventoried 665 statements, of which 657 were eligible, drawn from 46 of the 51 recordings across ten descriptive domains.
What we learned
In an authored storage comparison, ordinary narrative notes and the IHAI representation each retained all 657; an intrinsic storage advantage over ordinary notes was not demonstrated. The work shaped the design of the LifeCompass intake.
Limits
One clinic. Recordings are not confirmed to be distinct people. Source and storage judgments were unblinded and AI-only. Statements are what people reported, not verified diagnoses, and no clinical outcomes were studied.
Next
An independent, blinded reconstruction, and an audit of the extraction package.

WisdomMap

MKES (Medical Knowledge Extraction and Structuring) and the 知源中醫 knowledge project

Organizes health knowledge so that every statement can be traced back to where it came from.

  • Descriptive
  • 928 source occurrences preserved
  • 917 structured propositions
Explore WisdomMap
Who
Practitioners, researchers and stewards who rely on traditional and contemporary knowledge.
What
A traceable structure for health knowledge.
Why
Knowledge loses trust when its source is lost.
Where and when
Whenever knowledge is reused in teaching or care.
How
Rules, frozen before each run, merge statements only when they are confirmed equivalent; uncertain cases stay separate.
What we did
Tested MKES on a frozen corpus of 928 source occurrences of acupuncture knowledge. Results were compared with an earlier grouping only after each run's output was locked.
What we learned
In the refined second run, all 928 occurrences and their source links were preserved, producing 917 propositions with 6 unresolved exceptions and 0 final execution failures. That run reproduced 913 of 915 earlier occurrence groups exactly.
Limits
These figures measure reproducible structuring under fixed rules, not extraction accuracy, semantic truth or clinical validity. A planned six-channel analysis remains data-blocked. Findings from the 知源中醫 project have not yet been summarized for public release.
Next
The MKES methods are described in a manuscript prepared for peer-reviewed publication.

CareSignals

Relational-cue perturbation pilot

Tests whether AI notices when the timing or relationships in a health story change its meaning.

  • Exploratory
  • 29 constructed cases
  • 4 wording variants per case, across 3 models
Explore CareSignals
Who
Anyone whose health story is summarized by AI.
What
Whether AI interpretation responds to changes in relationships and timing.
Why
The same symptom can mean something different in a different context.
Where and when
Whenever AI reads a person's narrative.
How
Controlled wording changes to constructed cases, with each response scored.
What we did
Built 29 constructed cases, each in 4 wording variants, and scored how three AI models' interpretations responded to the changes: 348 recorded responses in all.
What we learned
Detailed results are being prepared for conference presentation and will be published here once disclosure requirements have been checked.
Limits
Constructed cases, not patients. Not a measure of clinical accuracy, not an independent replication, and no model is declared better.
Next
Publish the results after conference disclosure is cleared, then refine the control conditions.

RestWell

Paper 0: sleep and acupoint evidence review

Examines what existing studies of acupoint-based approaches to sleep can and cannot justify.

  • Descriptive
  • 26 effect calculations reproduced
  • 6 pooled clinical contexts
Explore RestWell
Who
People with sleep difficulties, and the practitioners who advise them.
What
The strength of published evidence for acupoint-based approaches to sleep.
Why
Practical advice should rest on what evidence can actually support, including where it is uncertain.
Where and when
Before any sleep-related support is offered within HCC.
How
Reconstruction of selected effect-size and pooled calculations from published trials.
What we did
Reconstructed selected sleep and acupoint evidence calculations, including null contrasts: 26 effect calculations and 6 pooled clinical contexts.
What we learned
Several key comparisons, including direct transcutaneous stimulation and laser contrasts, did not show a clear effect, and results varied widely between studies.
Limits
The arithmetic was verified from supplied extractions, not from the conduct of the original trials. Full study selection and risk-of-bias grading were not reconstructed. This is not evidence that HCC or any IHAI approach improves sleep.
Next
Informing the design of a planned sleep-focused study with an academic partner.

Human participation and sustainability

TimeTrust

Time-Trust contribution accounting

Explores how the time and care people give to their communities can be recorded and recognized.

  • Conceptual
Explore TimeTrust
Who
Stewards, volunteers, participants and partner organizations.
What
A record of human contribution to care.
Why
Much care work is unpaid and unseen, and what is unseen is hard to sustain.
Where and when
Each time someone gives time to support another person.
How
Contributors own their contribution records, IHAI keeps shared accounting, and exchanges are voluntary.
What we did
Designed the conceptual model. Logging the time that stewards and participants give is already part of how IHAI works.
What we learned
No empirical results yet.
Limits
Rules for accounting, verification and recognition are not yet specified, and no simulations or evaluations have been reported. TimeTrust is a research concept, not a financial product.
Next
Specify the rules, then evaluate them within a pilot.

What comes next

Planned

2027 implementation pilot

A pilot with approximately 20 participants will test whether another organization can run HCC from documented procedures, with consistent and auditable results. It will measure feasibility, acceptability, supervision and escalation fidelity, information quality, delivery costs and participant-reported self-care confidence. It is an implementation study, not a clinical trial. An application for ethics review is pending.

What we are not claiming

  • That HCC or any IHAI tool improves health outcomes. No clinical outcomes have been established.
  • That any IHAI tool can diagnose, treat or reliably predict a health condition.
  • That the prototype is validated for clinical decision-making. It is not.
  • That these results have been independently verified. They are our own research and testing.