Big data in health: uses, savings, disease management and privacy

Five uses, from the routine to the ambitious. The first is the one students overlook, and it is the largest.

• Knowing what is happening at all. Every statement in this module about trends, inequities, conditions and the health system rests on linked national data. Without it, health policy would be anecdote. This is not a future application; it is the foundation of the subject.
• Surveillance and outbreak response. Detecting a communicable disease outbreak, tracking immunisation coverage by area, and monitoring for medicine safety signals that only appear across very large numbers.
• Planning and evaluation. Where services are needed, whether a program worked, which hospitals have outcomes worth learning from, and where the gaps between groups are widening.
• Prediction and targeting. Identifying people at high risk of readmission, deterioration or a specific complication, so that limited attention can be directed where it will do most.
• Research and discovery. Testing whether a treatment works in the real world rather than only in a trial, finding rare harms, and identifying patterns across genetics, environment and behaviour that no single study could see.

The value of health data comes overwhelmingly from linkage. A prescribing record alone tells you what was dispensed; linked to hospital and mortality records it tells you whether the medicine helped or harmed. That is also exactly why linkage is the point at which privacy risk is greatest — the benefit and the risk have the same source.

The syllabus asks how big data is reducing healthcare spending. Answer it honestly, because the mechanisms are real and the savings are frequently overstated.

Where savings are genuine.

• Avoiding duplication. Knowing a test has already been done avoids repeating it — direct, immediate and measurable.
• Preventing avoidable admissions. Identifying people at high risk and intervening earlier is cheaper than the admission, when the follow-up is actually delivered.
• Identifying low-value care. Data shows procedures and tests that provide little benefit, and shows unwarranted variation — the same operation performed at very different rates in similar populations, which cannot all be clinically justified.
• Medicine safety and effectiveness in the real world. Finding harms and non-effects after approval avoids paying for treatments that do not help.
• Better targeting. Directing a screening or prevention program at the people most likely to benefit raises the health gained per dollar.

Where the claim is weakest.

• Prediction is not intervention. Knowing who will be readmitted saves nothing unless something is funded to change it. This is “data without a responder” again, and it is the most common reason a predictive program produces no saving.
• Finding more disease costs more. Data-driven detection frequently increases spending, appropriately, by finding people who need treatment. That is a health gain, not a saving.
• The infrastructure is expensive. Collecting, linking, governing and analysing data costs money and skilled people.
• Savings are often shifted rather than made. A cost avoided by one funder may reappear with another, which is the seam problem from Lesson 9 in a new form.

So the defensible version: big data improves the value obtained from health spending more reliably than it reduces the total. Given that spending growth is driven mainly by doing more that works, better value is the honest ambition.

• Finding causes. Linking exposure, genetic, environmental and outcome data across very large populations reveals risk and protective factors that no single study is large enough to detect. Rare effects and small effects only appear at scale.
• Precision medicine. Matching treatment to the characteristics of a person or a subgroup — most established in cancer, where the genetic profile of a tumour can determine which treatment will work. This is a direct application of large linked datasets to an individual decision.
• Faster and better research. Identifying eligible participants for trials, using existing records as comparison groups, and following outcomes long after a trial ends without asking participants to return.
• Real-world effectiveness and safety. A trial shows what happens in selected participants under ideal conditions; linked data shows what happens in everyone, including older people, people with several conditions, and people who do not take treatment exactly as prescribed — the people trials most often exclude.
• Managing chronic conditions. Registries, monitoring data and pathology results together allow a service to see which patients are drifting out of control before they present acutely.
• Quality improvement. Comparing outcomes between services reveals what better performers do differently, which is a mechanism for improvement that costs nothing new.

One caution to include, because it is the equity point in this lesson: a dataset that under-represents a population produces conclusions that fit that population worse. If research and prediction rest on data drawn mainly from metropolitan, English-speaking, well-served populations, the resulting treatments and models will be least accurate for the groups already worst served. That is the algorithmic bias argument from Lesson 16 arriving through the data rather than through the model, and it means representative data collection is an equity intervention, not an administrative detail.

The syllabus asks what measures need to be taken, so the answer is a set of protections with reasons, not a general statement that privacy is important.

• Collect the minimum, and keep it only as long as it is needed. The strongest protection for data is not holding it. Every additional field and every extra year is additional risk with no benefit unless it is used.
• De-identify, and treat de-identification as imperfect. Remove names and identifiers, and recognise that combinations of age, postcode, sex and dates can re-identify a person, particularly where a condition is rare or a population small. De-identification reduces risk; it does not eliminate it, and claiming it does is the most common error in this area.
• Control access by purpose, not by person. Access should be granted for a defined use, limited to the fields needed, time-limited, and logged. “Trusted staff” is not an access control.
• Independent oversight and approval. An external body, not the organisation that wants the data, should decide whether a secondary use is justified.
• Transparency and meaningful choice. People should be able to find out what is held, who has looked at it, and how to restrict it — and the choice must be real, not buried.
• Technical security. Encryption in storage and transit, authentication, monitoring, and tested response plans, because health data is a target and breaches are common.
• Enforceable consequences. Penalties for misuse and unauthorised access that are actually applied. A rule with no consequence is a suggestion.
• Analyse without moving the data where possible. Sending the question to the data, rather than copying the data to the researcher, removes an entire class of risk.

The trade-off is real and cannot be argued away. Stronger privacy protection reduces what can be learned; weaker protection increases what can be learned and the harm when it goes wrong. The task is not to eliminate the tension but to place the line deliberately and defend it — and to remember that a serious breach costs the public trust on which all future use depends.

The dot point says shaping, which is a stronger word than helping. Four ways it shapes rather than merely assists.

• It decides what is visible. A problem that is measured becomes a policy problem; one that is not stays invisible. Data collection is therefore a political act as much as a technical one, and choosing to disaggregate by group is how an inequity becomes something a system has to answer for.
• It shifts care from reactive to anticipatory. Prediction allows attention before an event rather than after, which is a change in what care is, not only in how it is delivered.
• It changes what individual treatment means. Precision medicine substitutes group-average prescribing with treatment matched to a person's characteristics.
• It redistributes power over information. Whoever holds and can analyse the data holds the ability to define problems and evaluate solutions. Where that is public institutions with oversight, it serves the population; where it is commercial holders, the questions asked are the questions worth money to answer.

Which produces the judgement this dot point is looking for: big data is one of the largest opportunities in health and it is not automatically a benefit. Its value depends on representative collection, on something being funded to act on what is learned, on governance the public can trust, and on deliberate attention to who benefits — because the distribution rule applies here as it does to every other technology in this inquiry question.

For each proposed use of linked health data, name the benefit, the specific privacy risk, and the one protection you would insist on.

A proposal would link every Australian's hospital, prescribing and pathology records into one research dataset, available to approved researchers, on an opt-out basis.

Write your position. State what you would allow, what you would refuse, and the conditions you would require. Then do the part that most answers skip: say what is lost by your own conditions — what could not be learned because of the line you drew — and why you accept that loss.