Why most of the data in a medical record is never reused

A routine check-up produces data. A blood pressure reading, a note on chest pain, perhaps a prescription and a lab order. It all goes into the electronic health record, where it serves the next consultation. What usually does not happen is that it gets compared with the patient's earlier history in a systematic way, set against thousands of similar patients, or used to spot a problem before it becomes visible. A Tech Anatomy episode published in April 2025 asked why.

A medical record is richer than most people assume. It holds clinicians' free-text notes, prescriptions, laboratory results, imaging, diagnoses coded for billing and, increasingly, data from connected devices. The episode identified three kinds of obstacle to using it more fully: practical, legal and economic.

Practical obstacles

The first is fragmentation. A general practitioner, a hospital and a specialist often work in different software that does not exchange data, so no single place holds the full picture. The second is the shape of the data itself. Much of the clinically important information sits in free-text notes and images rather than structured fields, and the same concept may be recorded in different ways across sites. The third is time. No clinician reads a patient's full history line by line during a consultation, and analysing it at scale requires data preparation that most care settings are not staffed for.

Legal obstacles

Health data is among the most protected categories of personal information. In the European Union, the General Data Protection Regulation classes it as a special category whose processing is prohibited unless one of a limited set of conditions applies. A hospital that wants to train a model to predict, say, which patients are likely to return to the emergency department cannot simply send its records to a research team or a company.

Removing names is not enough to make data anonymous. Under the GDPR, data that can still be linked back to a person by reasonably likely means remains personal data. Research has shown how low that bar can be: a 2019 study in Nature Communications estimated that 99.98 percent of Americans could be correctly re-identified in any dataset using 15 demographic attributes such as age, sex and postcode. Medical images add their own risks: in a 2019 study in the New England Journal of Medicine, face-recognition software matched reconstructions of research participants' head MRI scans to their photographs in most cases.

Economic obstacles

The episode pointed to a quieter barrier: uncertainty about value. A hospital that shares its data takes on legal and reputational risk, while the benefit, a better model, may accrue to a company or to other institutions. Who owns the insights, and who is paid for them, is often unresolved. Until incentives are clearer, the cautious choice is not to share.

Two responses being tested

Synthetic data is the first. A generative model learns the statistical patterns of a real dataset, then produces artificial patient records that reproduce those patterns without corresponding to any real person. In principle, such data can be shared more freely, used to develop and test software, or used to fill gaps in a dataset. A 2023 narrative review in PLOS Digital Health identified seven main uses, from research and public health to software development and training. It also listed the limits: a recognised risk that information about real patients leaks into the synthetic records, dependence on the quality of the model that generated them, and the fact that synthetic data does not always reproduce the properties of the original precisely. Its authors concluded that real data remains the preferred choice, with synthetic data a way to bridge access gaps. The episode's statement that synthetic data can be "just as useful as real data" is therefore true for some tasks and not as a general rule. How synthetic data is generated and evaluated is covered in a separate Tech Anatomy article on synthetic health data.

Digital twins are the second. Here a patient's own data is used to build a computational model that can be queried. At Johns Hopkins, researchers built heart models from patients' MRI scans to plan where to ablate tissue in a rhythm disorder, in a 2018 proof-of-concept study. In type 2 diabetes, the company Twin Health has run a randomised trial in India of a programme that uses sensor data to personalise diet and lifestyle advice; interim results presented in 2022 reported high remission rates, in a company-funded study whose final peer-reviewed results remain to be seen. Tech Anatomy has a separate article on digital twins in health.

Competing readings

One reading treats the under-use of health data as a failure with a human cost: patterns that could prevent disease remain invisible, and AI tools are trained on narrow datasets because broader ones are locked away. On this view, synthetic data, twins and new legal frameworks are overdue.

Another reading sees much of the caution as justified. Health data breaches harm people, consent for secondary use is often weak, and techniques presented as privacy-safe have been shown to leak information. On this view, slow and controlled access is the price of public trust, without which patients may stop sharing information with their doctors at all.

What the evidence does not show

There is little direct measurement of how much clinical benefit is lost because data goes unused; the argument rests mostly on plausible examples. The privacy guarantees of synthetic data depend on how it was generated, and evaluations are not standardised. And the published clinical evidence for patient-specific twins is still dominated by small or early studies.

What to watch

In Europe, the main test will be the European Health Data Space, adopted in 2025, which creates national bodies to grant access to health data for research and innovation under defined conditions, with obligations phasing in from 2027. Whether those bodies process requests quickly enough to matter, and whether synthetic data gains an accepted status within that framework, will say a lot about whether the obstacles described here are being lowered or only reorganised.

Listen to the episode

Why Doctors Don’t Use All Our Data

Listen · 8 min