The Health Shadow Ledger

Imagine a health system that could tell us how many illnesses were treated last year. Now imagine one that could also estimate how many might never have occurred if an early warning had been recognised, a diagnosis had arrived sooner, a home had been warm enough, or a person had been able to obtain help before their situation became a crisis.

We have increasingly sophisticated ways to count hospital admissions, diagnose disease, measure waiting lists and record deaths. Yet a fundamental question remains difficult to answer: how much human suffering occurs because the right information, the right intervention or the right opportunity failed to meet the right person at the right time?

This is not simply a question about clinical performance. It reaches into housing, employment, education, transport, social isolation, environmental exposure, digital access and the organisation of public services.

It also raises a potentially transformative question for artificial intelligence: could AI help society build a shared intelligence system for preventable harm, rather than merely automate parts of the healthcare system that already exists?

This article proposes a concept called the Health Shadow Ledger. It is not an existing national database, a clinically validated algorithm or a claim that every illness can be predicted. It is a proposed public-interest framework for connecting fragmented evidence, identifying patterns of preventable harm and helping communities test which interventions genuinely improve health.

1. The blind spot: every institution sees a fragment

A general practitioner sees one part of a person’s health. A hospital sees another. A housing department may know about damp or repeated repair requests. A school may notice persistent absence. A pharmacy sees medication collection. A local authority may hold information about fuel poverty, transport or social care needs.

Each record can be useful on its own. But the public-health significance may sit between the records.

Consider a hypothetical person whose asthma repeatedly worsens. Their clinical record shows recurring symptoms. Their housing record contains unresolved damp complaints. Their employment circumstances make appointments difficult. Their energy costs encourage them to limit heating and ventilation. Each service may process its own task without any single organisation recognising the full pattern.

That example is illustrative, not a claim about a specific patient or a quantified outcome. It demonstrates a structural problem: the information needed to understand a health outcome can be distributed across systems that were never designed to learn together.

Public-health researchers already investigate social determinants of health and inequalities in access and outcomes. The proposal here is not that those disciplines have overlooked the issue. It is that society could develop a more systematic way to turn fragmented evidence into a continuing, testable account of preventable harm.

Research published in PLOS Digital Health in January 2026 examines how missing or decontextualised healthcare data can conceal the social circumstances that shape what gets recorded. Its argument reinforces an important principle: incomplete data is not always a neutral technical inconvenience. It may reflect how institutions work, whose experiences are captured and whose are not.

The question is therefore bigger than whether AI can read more records. It is whether we can build an accountable system that understands the limitations of those records and helps people act on the patterns they reveal.

2. What exactly is a Health Shadow Ledger?

The proposed ledger would be a public-interest intelligence framework, not a single giant database containing everyone’s medical history. It would seek to answer five questions:

The ledger would link evidence to questions, not automatically link every detail of every person’s life into a permanent profile. Where possible, analysis could use aggregated statistics, privacy-preserving methods, approved research environments and information that remains with the organisations legally responsible for it.

Its distinctive feature would be a structured account of the pathway from a potential risk to a measured outcome. A signal would not be treated as proof. A correlation would not be treated as causation. A model’s prediction would not be treated as a medical diagnosis.

Each finding would carry its source, date, confidence, known limitations and the evidence required to test it. Where evidence is weak, the ledger should display uncertainty rather than manufacture precision.

3. What AI could do that conventional reporting struggles to do

AI is not valuable simply because it can process enormous quantities of information. Its potential public-health value lies in finding connections, helping humans examine competing explanations and making complex evidence more usable.

A. Detect patterns across disconnected systems

With lawful access to appropriately governed data, analytical models could help researchers examine whether certain combinations of housing conditions, service delays, clinical histories and environmental exposures are associated with worsening outcomes.

Natural-language processing could also help organise unstructured information, such as free-text reports or complaints, into themes for investigation. Researchers would still need to validate the classifications, examine missing data and test whether apparent patterns survive outside the dataset in which they were discovered.

B. Reconstruct the timeline before a crisis

Instead of analysing only the final hospital admission, an AI-assisted review could help authorised investigators reconstruct the sequence of events leading up to it.

Were there earlier contacts? Did symptoms recur? Was a referral delayed? Did a service fail to respond? Were the apparent warning signs actually visible at the time, or do they only look obvious in hindsight?

That final question matters. A system must not condemn clinicians or patients for failing to act on information that was unavailable, ambiguous or impossible to interpret at the time. Retrospective pattern recognition can create an illusion that an outcome was predictable when it was not.

C. Identify neglected questions, not just predicted patients

Many predictive systems are designed to estimate which person is most likely to experience a particular outcome. The ledger would add a different task: identifying where the health system lacks the evidence to answer a consequential question.

For example, AI might flag that a community has a high burden of respiratory illness but inadequate information about housing conditions, or that an intervention is routinely evaluated for average outcomes but rarely for people facing several overlapping barriers to care.

The next step would not automatically be more surveillance. It could be better research, a community consultation, an environmental investigation or a properly designed service evaluation.

D. Help test possible solutions before scaling them

AI could help compare potential interventions, identify relevant research and estimate where further evaluation might have the greatest value. For instance, a local team could investigate whether a combined housing-and-health referral pathway merits a controlled pilot.

But a model’s estimate is not evidence that an intervention will work in practice. Promising ideas require appropriate study designs, clinical and community input, cost evaluation and monitoring for unintended effects. Where feasible, randomised or strong quasi-experimental evaluations can help distinguish a real effect from changes that would have happened anyway.

E. Translate evidence into understandable choices

Generative AI could help turn complex research into accessible summaries for patients, clinicians, community groups and policymakers. It could explain what is known, what remains uncertain and which questions need further investigation.

It must also show its workings in a practical sense: cite the underlying evidence, distinguish quotation from interpretation, flag conflicting findings and allow a qualified human to check the conclusion. A fluent answer is not necessarily a correct one.

4. The most important discovery may be what the data cannot tell us

Here is the central intellectual challenge. Health data is not a transparent window onto reality. It is a record of what particular institutions were able and willing to observe.

People who cannot obtain an appointment may leave fewer clinical records. People who struggle with digital forms may be less visible in online services. A symptom that is dismissed may never become a diagnosis. A person who avoids seeking help because of previous treatment may appear, in a dataset, to have fewer needs than someone who repeatedly attends.

These are examples of mechanisms that researchers should investigate, not assumptions that apply to every person or service. Their importance is that the absence of a recorded event can have several explanations. It cannot safely be interpreted as proof that the event, need or barrier did not exist.

In a 2025 paper in The Lancet Digital Health, the STANDING Together consensus recommendations addressed bias and transparency in health datasets, including the need to describe dataset limitations and assess implications for different population groups.

That research points towards a crucial design principle for the ledger: every important conclusion should be accompanied by an account of who may be missing from the evidence.

That account should examine at least four possibilities:

• Coverage: who is not represented in the available data?
• Participation: who could not, would not or was never invited to contribute?
• Interpretation: whose symptoms, circumstances or experiences may have been misclassified?
• Power: who decides which outcomes matter, which questions are funded and which errors are considered acceptable?

An AI system that predicts accurately for the people already well represented in its training data can still fail those who are least visible. A system that learns institutional habits without questioning them may reproduce those habits at scale.

The ambition should therefore be more than a larger dataset. It should be a better account of the boundaries of our knowledge.

5. From predicting illness to measuring preventable harm

Prediction is only one stage of prevention. A model may identify a risk without revealing its cause, and a correct prediction does not guarantee that an effective response is available.

For the Health Shadow Ledger to be useful, it would need to organise its findings around a chain of evidence:

This framework could be applied to a range of public-health questions, subject to evidence and feasibility.

Respiratory health: Are recurring symptoms associated with remediable environmental exposures, and does a targeted intervention reduce exacerbations?

Medication safety: Can fragmented records reveal recurring risks around transitions between hospital, primary care and home, and can a tested process reduce avoidable harm?

Earlier diagnosis: Are particular groups experiencing longer diagnostic journeys, and which specific changes reduce the delay without causing excessive false alarms?

Maternal health: Are warning signs, access barriers or differences in care associated with adverse outcomes, and do targeted improvements narrow those differences?

Mental health and isolation: Can communities evaluate supportive, voluntary interventions that improve access to help without turning ordinary social behaviour into a surveillance target?

In every case, the aim is to test a modifiable cause or service failure, not to label people as problems to be managed.

6. The economic question: what does preventable suffering cost?

Healthcare budgets record expenditure. Economic evaluations can estimate the costs and benefits of interventions. But the wider consequences of preventable illness can be distributed across households, employers, schools, social care and public services.

A condition that could have been prevented or treated earlier may affect a person’s ability to work, care for family, study or live independently. It may also impose financial and emotional burdens on people close to them. These effects should not be reduced to a single monetary value, but ignoring them can make the overall consequences of illness harder to understand.

The proposed ledger could help researchers map these consequences and identify where the costs of a problem fall compared with the organisation that could help prevent it.

For example, an intervention may require an initial investment from a housing or local-government service while potential benefits appear later in health outcomes, reduced service use or improved quality of life. Whether the intervention saves money overall is an empirical question, not something that should be assumed.

This creates a difficult but important policy problem: the institution expected to pay for prevention may not be the institution that receives the benefit.

A shared evidence framework could help policymakers examine that mismatch. It could compare intervention costs, health outcomes, distributional effects and the timing of benefits, while making assumptions visible. It could also reveal when a proposed intervention improves health but does not generate a net financial saving.

That distinction matters. The purpose of public health is not simply to minimise expenditure. It is to improve health fairly, use resources responsibly and prevent avoidable suffering where credible evidence supports doing so.

7. The dangers: a ledger could become a surveillance machine

A system designed to make hidden harm visible could itself create harm if it is built without strict limits.

Linking health, housing, financial or behavioural information can expose intimate details. A risk score can become a label. A prediction can influence decisions about services, insurance, employment or eligibility. A model can confuse poverty with risk, lack of recorded data with lack of need, or a population-level association with an individual’s future.

Even a system created with good intentions could be repurposed. Public trust would be impossible to sustain if people believed that asking for help might make them more likely to be monitored, penalised or excluded.

The World Health Organization’s guidance on the ethics and governance of large multimodal AI models in health discusses the potential of these technologies alongside concerns about accuracy, bias, privacy and accountability.

The Health Shadow Ledger would need safeguards that are part of its architecture, not promises added after deployment.

Privacy-preserving technology is not a substitute for governance. Nor is public benefit a blanket justification for collecting sensitive information. Every proposed use must be assessed for necessity, proportionality, legality and the risk of misuse.

Some questions should not be answered through individual-level data at all. A population-level problem may be better investigated through environmental monitoring, anonymous surveys, community research or improved service statistics.

8. How PHAT could turn this idea into a real research programme

The proposal does not need to begin with a national platform or a vast new database. In fact, starting smaller would make it easier to test whether the concept creates genuine public value.

PHAT could help frame the public questions, bring evidence together and convene clinicians, public-health researchers, data specialists, local authorities and community representatives around a bounded pilot. Organisations such as Made2Master could be mentioned in future work where there is a relevant, clearly defined role, but participation or endorsement should never be implied without confirmation.

A credible first phase could follow six steps:

1. Choose one measurable problem. Select a specific outcome with credible evidence of potential preventability, rather than attempting to model every health problem at once.
2. Map the evidence before collecting more data. Identify existing datasets, research, known limitations, data owners and legal constraints. Document whose experiences are missing.
3. Bring affected people into the design. Ask residents, patients and frontline professionals which failures matter and what forms of analysis they would find useful or unacceptable.
4. Establish a baseline. Define the outcome, comparison method, costs, subgroup measures and evaluation period before deploying a model or intervention.
5. Test a narrow AI-assisted workflow. Use the minimum necessary data to identify and investigate patterns. Require human review and document errors, uncertainty and performance differences.
6. Publish the result, including failure. Independent evaluation should determine whether the approach improves decisions or outcomes enough to justify its costs and risks.

The initial product could be a public report rather than software. It could document a preventable-harm pathway, the strength of the evidence, the information gaps, the possible interventions and the questions a future pilot must answer.

This would establish whether the underlying framework is useful before investing in complex technology. It would also make the idea easier for researchers, funders and public institutions to scrutinise.

9. What success should look like

A ledger that generates impressive predictions but changes nothing should not be considered a success. Nor should a system be rewarded simply for identifying more people as high risk.

Its performance should be judged by whether it supports better decisions and measurable improvements in health, with acceptable costs and risks.

These are proposed evaluation dimensions, not a validated scoring instrument. Their measures and relative importance would need to be agreed for each pilot before results are known.

10. The deeper idea: prevention is a form of collective intelligence

Modern societies have developed remarkable systems for detecting events after they happen. We record admissions, investigate outbreaks, publish mortality statistics and monitor service performance. These capabilities are essential.

The next challenge is to connect what we learn from those events to a disciplined process of prevention.

Imagine if every serious public-health investigation could contribute to a growing body of structured knowledge: what warning signs were present, what evidence was available, which intervention was attempted, what changed and which groups benefited. Imagine if negative results were retained rather than quietly forgotten, and if communities could see the reasoning behind decisions that affected them.

AI could help organise this knowledge, compare new situations with past evidence, expose contradictions and identify questions worth investigating. Humans would still need to decide which goals are legitimate, what risks are acceptable and how scarce resources should be distributed.

In that sense, the most important advance may not be a machine that claims to know which person will become ill. It may be an institution that becomes better at learning from the conditions that repeatedly make people ill.

This would require a change in how we think about intelligence. Intelligence would not mean merely producing a prediction. It would mean improving the connection between observation, explanation, intervention and learning.

And it would require a change in how we think about responsibility. A preventable outcome may not be attributable to one clinician, one department or one decision. Sometimes it emerges from a chain of individually understandable decisions that collectively produce a poor result. A carefully designed ledger could help investigate that chain without turning every finding into a hunt for someone to blame.

11. The question we should ask now

There is an understandable temptation to measure the future of AI in healthcare by counting automated tasks, faster reports, more accurate image analysis or reduced administrative time. These may all deliver value when properly demonstrated.

But a society can become more technologically advanced without becoming equally effective at preventing suffering. Efficiency in an existing process does not necessarily tell us whether the process addresses the most important problem, reaches the right people or tackles the causes of poor health.

The Health Shadow Ledger offers a different question for researchers, policymakers, innovators and public institutions:

We do not yet know whether a national or regional ledger built on these principles would deliver benefits at scale. Its feasibility, legal basis, cost, data requirements and effects would all need to be established through research and carefully governed pilots.

That uncertainty is not a reason to abandon the question. It is a reason to formulate it precisely, investigate it honestly and test the answer in the real world.

AI may eventually help medicine discover new treatments and recognise disease earlier. It may also help society understand the preventable conditions surrounding disease, learn from fragmented evidence and evaluate whether public interventions make a meaningful difference.

The ambition should be neither blind faith in technology nor resistance to it. It should be to make technology answerable to a clearer public purpose.

Perhaps one of the most valuable measures of a health system will not be how much information it collects, how many predictions it generates or how advanced its software becomes. It will be whether the system gets better at turning what it knows into fewer preventable crises and better lives.

That is the question the Health Shadow Ledger is designed to put on the table.