Which factors work together?
Some factors strengthen each other and some cancel each other out. We find those pairs and show them as a map.
Acchorda Research Collective · London
Know why.
Know what would change.
We study how health AI reaches its decisions, which factors work together inside it, and what realistic change would lead to a different outcome.
- Explainable AI
- Factors that interact
- Decision margins
- Realistic next steps
- Public health
- Clinical risk models
- Open research
What we study
Three questions we ask of every health AI model
Most explanations stop at "this factor mattered". We go further and ask how factors work together, how close a decision is to changing, and whether the suggested change is realistic.
How close is the decision?
We measure the gap between the model's top answer and its nearest rival. The smaller the gap, the easier the decision is to change.
Is the next step realistic?
A suggestion only helps if a real person could achieve it. We keep every suggestion inside sensible, agreed limits.
Our approach
CHORD: explanations built from how factors combine
CHORD is the method behind our work. Like notes in a musical chord, factors in a model can sound together, strengthen one another or soften one another. CHORD listens for those combinations and uses them to explain a decision and to find the smallest realistic change that would alter it.
- See the reasonsWhich factors matter most, and which ones pair up.
- Realistic next stepsSuggestions stay within safe, achievable limits.
- People stay in chargeAI supports professional judgement. It never replaces it.
How we work together
From first conversation to shared results
Step 1
Talk to us
Tell us about your service, your data or the question you want answered.
Step 2
Explore together
We look at what you have and agree what would be most useful to find out.
Step 3
Build and test
We apply our approach to your setting and check the results with your team.
Step 4
Share what we learn
Reports, joint papers, grant bids or a shared pilot.
3peer-reviewed conference papers in 2026: IEEE CAI, FTC and IEEE SIME
5areas of health studied so far, including heart health, diabetes, pregnancy and thyroid
1shared method, CHORD, extended step by step across every paper
Who we work with
Built for people who care for others
Public health teams
Local authorities, health agencies and community programmes.
Hospitals and clinics
Clinical teams using, or planning to use, risk prediction tools.
GP practices and primary care
Practices and primary care networks working on prevention.
Charities and community groups
Organisations supporting people's health and wellbeing locally.
Universities and researchers
Joint studies, student projects and grant partnerships.
Health-tech companies
Teams building digital health tools who want them to be trusted.
How it works
CHORD in six small steps
No maths needed. Each step below has a small live demo you can play with. The examples use simple made-up numbers unless a step says it comes from one of our papers.
1The decision margin
Every prediction is a close race or a clear win
When a model sorts someone into low, medium or high risk, it gives each option a score and picks the highest. We look at the gap between the winner and the runner-up. We call that gap the decision margin.
A large margin means the decision is settled. A small margin means a small change could tip it. When the margin reaches zero, the prediction changes.
Move the slider to change one factor and watch the race.
Low riskmargin 0.00
Illustration with made-up scores
2The main driver
Find the factor the decision leans on most
Next, we gently test each factor one at a time and see how much the margin moves. The factor that moves it most is the main driver for that person.
This is the familiar part of explainable AI. Most methods stop here. CHORD is just getting started.
Illustration with made-up values
3The partner factor
Factors rarely act alone
Then we ask a second question: which other factor changes how strongly the main driver works? We call it the partner factor.
Like two notes in a chord, partners can reinforce each other, so together they push harder, or dampen each other, so together they push less. Knowing which is which explains why the same factor matters more for one person than another.
Main driverPartner factorTogether: stronger
4The nudge test
Nudge the partner first, and the main change gets smaller
Here is the idea at the heart of CHORD. If we first move the partner factor by a tiny amount, 1% or 5%, how much does the main driver still need to change to tip the decision?
When the two reinforce each other, the answer can drop sharply. Pick a nudge to see a real result from our thyroid study.
0% less change needed than with no nudge
5Keep it realistic
A suggestion only counts if a real person could reach it
Maths alone can suggest changes no one could make: a weight far outside the human range, or a younger age. So we mark some factors as fixed, such as age, and give the rest realistic limits agreed in advance.
Any suggestion that leaves those limits is set aside, and we report how many were lost. We call that share the plausibility gap.
6% to 83%of suggested changes fell outside realistic limits, depending on the factor (heart and diabetes models)
29% to 75%of people received a realistic suggestion from our method, against at most 8% for a common alternative
6The interaction map
From one person to the whole picture
Finally, we add up the partner relationships across everyone in a test group and draw them as a map. Each line joins two factors the model treats as a pair. Red lines reinforce, blue lines dampen, and thicker lines are stronger.
The map shows what the model has learned. It describes the model's behaviour, not proven biology, and that distinction matters when people read it.
Reading our results fairly
What our findings mean, and what they don't
About the model, not the body
Every pairing we report is something a trained model has learned from data. It can point to useful questions, but it is not a confirmed medical mechanism.
Smaller steps, fewer answers
Our suggestions change far less than a leading alternative, usually one or two factors. The trade-off is that we find a suggestion for fewer people.
Support, not a substitute
Nothing here is medical advice. Our work is designed to help professionals ask better questions, with people firmly in charge of decisions.
Our research
One method, growing paper by paper
Each paper adds one idea to CHORD. Read them in order to see how the method grew from a yes-or-no decision to whole maps of how factors work together.
CHORD: guiding counterfactuals with higher-order sensitivity
The first paper introduced CHORD for yes-or-no decisions. It finds the main driver and its partner factor, then shows that nudging the partner first can reduce the change needed in the main driver to flip a prediction.
- Tested on two well-known public datasets, UCI Adult Income and Bank Marketing
- Introduced the nudge test at 1% and 5%
Margin-guided counterfactuals
Real health decisions often have three or more levels, such as low, medium and high risk. This paper adds the decision margin, which turns any number of options into one simple gap, plus a targeted margin aimed at a particular outcome of concern.
- Maternal health risk, fetal health and thyroid disease datasets
- Biggest effect on thyroid: a 5% nudge more than halved the change needed
- Changes about two factors, against about five for a leading alternative, with less overall movement
Interaction networks and plausibility
This paper moves from single people to whole groups. It draws interaction maps of what a model has learned and checks every suggestion against realistic limits, reporting how many are lost.
- Heart (hypertension) and diabetes risk models
- Realistic suggestions for 29% to 75% of people, against at most 8% for a common alternative
- Interaction rankings agreed with an independent method
Where we are heading
These directions are early and exploratory. We are looking for collaborators on each of them.
- Weight and diabetes medicines. Studying interaction maps for people receiving GLP-1 therapies, using national health survey data.
- More health areas. Testing the method on larger and more varied populations, so findings travel beyond a single dataset.
- Tools people can use. Turning the maps and suggestions into clear views that clinicians and public health teams can question.
Papers
Read our work
All three papers behind CHORD, newest first. Open one to read it here, page by page, or copy the citation for your own work.
Interaction Networks for Cardiometabolic Risk Models: Auditing Learned Model Structure and Clinical Plausibility Across Chronic Disease Domains
In short: draws maps of which factors a heart or diabetes risk model treats as pairs, and checks whether its suggested changes are realistic. Many standard suggestions turn out not to be.
Cite this paper
A. Fatima and Z. Shahid, "Interaction Networks for Cardiometabolic Risk Models: Auditing Learned Model Structure and Clinical Plausibility Across Chronic Disease Domains," IEEE SIME, 2026.
Margin-Guided Counterfactuals: Extending Higher-Order Sensitivity Analysis to Multi-Class Clinical Risk Stratification
In short: extends CHORD from yes-or-no decisions to three or more risk levels using one simple gap, the decision margin. Tested on maternal health, fetal health and thyroid data.
Cite this paper
A. Fatima and Z. Shahid, "Margin-Guided Counterfactuals: Extending Higher-Order Sensitivity Analysis to Multi-Class Clinical Risk Stratification," in Proceedings of the Future Technologies Conference (FTC), 2026.
CHORD: A Novel XAI Framework for Guiding Counterfactuals via Higher-Order Sensitivity Analysis
In short: the paper that introduced CHORD, the main driver, the partner factor and the nudge test, for yes-or-no decisions. Tested on the UCI Adult Income and Bank Marketing datasets. Presented in Granada, Spain, in May 2026.
Cite this paper
Z. Shahid, Y. Rahulamathavan and S. Dogan, "CHORD: A Novel XAI Framework for Guiding Counterfactuals via Higher-Order Sensitivity Analysis," in 2026 IEEE Conference on Artificial Intelligence (CAI), IEEE, 2026, pp. 1817-1822.
Papers are shared as the authors' versions for study. Please cite the published versions.
Founders
Public health and AI, side by side
Acchorda brings together a public health specialist and an AI researcher, so everything we build starts from real people and real services.
Founder
Amber Fatima
Public Health Specialist · Lecturer
- Department of Public Health and Social Care, QA Higher Education, London
- London Metropolitan University
Amber founded Acchorda to bring a public health view to health AI. She leads the collective's research direction and partnerships, keeping the work focused on questions that matter to communities, patients and health services.
Co-founder
Zohaib Shahid
PhD Scholar · Lecturer
- Institute for Digital Technologies, Loughborough University London
Zohaib developed CHORD, the AI method at the heart of Acchorda's work. He leads the technical side, from building and testing models to turning the method into tools people can use.
Collaborate
Let's build something useful together
Whether you have data, a service to improve or a research idea, we would like to hear from you. Research first; collaborators welcome.
Research partnershipsPilot projectsService evaluationGrant applicationsTeaching & workshopsCo-authored papers