MyTwin
Back to the blog
Patients5 min read

Predictive health: anticipating without claiming to predict the future

Predictive health does not read the future. It uses data to estimate the probability of an event or a change within a defined context. Well understood, it can support prevention. Poorly presented, it can create false certainties. Here are the key points for interpreting a risk score.

By Rubens Valcy

Founder of MyTwin

Published on

Contents
  1. Prevention, screening and prediction: three different concepts
  2. How is a predictive model built?
  3. Five questions to ask about a risk score
  4. The main limits of predictive health
  5. From prediction to shared decision-making
  6. Frequently asked questions
  7. Sources

Predictive health refers to the use of data and models to estimate the probability of a health event, a change in health status or a response to an intervention. The result can take the form of a score, a risk category or a trend curve. It is not an individual prophecy, but an estimate calculated from comparable people or situations.

This distinction is fundamental. A 10% risk does not mean the event will happen, and a low risk does not mean it is impossible. The value of an estimate depends as much on the quality of the model as on how well it guides a useful action.

An estimated 10% risk

Out of 100 people with a comparable profile, about 10 would experience the event.

The score does not say which ones.

Prevention, screening and prediction: three different concepts

Prevention
Covers the actions intended to prevent a disease, reduce a risk or limit the consequences of a problem that is already present.
Screening
Looks for a disease or an abnormality in people who do not necessarily have symptoms.
Prediction
Estimates a future probability, or the likely presence of a condition, from observed variables.

These approaches can complement one another. A score can help identify the people for whom a check-up or a medical conversation is particularly relevant. But a score does not replace the reference test, the clinical examination or medical reasoning.

How is a predictive model built?

Developers start by defining an outcome to predict: hospitalization, complication, recurrence, response to treatment or the evolution of a biomarker. They then select variables that are available before that outcome, such as age, certain medical history, lab results or repeated measurements.

The model learns or estimates relationships from a dataset. It must then be evaluated on other data. Good performance in the medical center, country or population used to develop it does not guarantee the same performance elsewhere.

The TRIPOD+AI statement, published in The BMJ in 2024, sets out 27 items to make studies of prediction models using regression or machine learning transparent. PROBAST+AI, published in 2025, helps assess a model’s quality, risk of bias and applicability. These frameworks are a reminder that an accuracy percentage on its own is not enough to judge a tool.

Five questions to ask about a risk score

What exactly does the score predict?

The outcome must be clearly defined, along with the time period. “Cardiovascular risk” is too vague if it does not specify which event is considered and over what time frame.

In which population was it evaluated?

A model can be less reliable for profiles that are underrepresented in the data. The World Health Organization points out that training datasets may exclude or underrepresent certain populations, which can reproduce inequalities.

How well is it calibrated?

Discrimination separates people at higher and lower risk. Calibration checks that the predicted probabilities match the frequencies actually observed. Both are necessary.

What happens after the result?

A score is useful if it leads to a proportionate action: rechecking a measurement, seeking advice, adjusting follow-up or applying a validated recommendation. Without a clear pathway, it can only cause worry.

Who interprets the information?

A professional can take into account the person’s symptoms, medical history, treatments and preferences. An isolated automated reading lacks this context.

The main limits of predictive health

A model can produce false positives, which flag a risk with no corresponding event, and false negatives, which wrongly reassure. Its performance can also decline when practices, devices or the population change. This is sometimes called model drift or data drift.

There is also a risk of overdiagnosis or overmedicalization if every variation triggers a series of tests with no proven benefit. Conversely, a well-integrated estimate can help prioritize prevention or a check-up.

From prediction to shared decision-making

The best use of a predictive tool is usually a structured conversation. The result can explain why an action is being suggested, but the decision must take into account the benefits, risks, alternatives and the patient’s preferences.

MyTwin for patients offers a journey that combines data, prevention tools and conversations with a healthcare professional. This combination is essential: digital results should support understanding and follow-up, not replace clinical judgment. Clinicians can explore the framework we offer on MyTwin for clinicians. Our privacy policy sets out the framework used for personal data.

Frequently asked questions

Sources

  1. Collins GS et al., 2024, “TRIPOD+AI statement”, BMJ, 385:e078378.
  2. Moons KGM et al., 2025, “PROBAST+AI”, BMJ, 388:e082505.
  3. World Health Organization, June 28, 2021, “Ethics and governance of artificial intelligence for health”.
  4. Haute Autorité de santé (France’s National Authority for Health), April 9, 2025, “Numérique et intelligence artificielle à la HAS”.

This article is provided for information purposes only. It does not replace advice, diagnosis or treatment from a healthcare professional.

Keep reading