A new AI model can forecast a person’s risk of diseases across their life

Foreseeing Health: Delphi-2M and the Next Generation of Disease Prediction

Advances in artificial intelligence (AI) are increasingly reshaping healthcare. A recent development—the model called Delphi-2M—represents a major step in using AI to forecast a person’s risk of many diseases well in advance, potentially over decades. Developed by teams at the European Molecular Biology Laboratory (EMBL), the German Cancer Research Center, and the University of Copenhagen, Delphi-2M moves beyond traditional single‐disease risk calculators, aiming instead for a holistic, long‐term view of health.

How It Works

Delphi-2M was trained on large, anonymised longitudinal health datasets: medical records from about 400,000 UK Biobank participants, and validation on 1.9 million people from Denmark’s national patient registry.

The data inputs include:

• Demographic factors (age, sex)
• Medical history — prior diagnoses, timing of diseases, order of events in health records
• Lifestyle factors — such as smoking, alcohol use, obesity/BMI, perhaps other behavior‐indicators

Delphi-2M uses algorithmic architectures similar to generative AI / large language models, but adapted for health‐trajectory prediction rather than text generation. It treats medical events like “tokens” in a sequence: not just whether someone had disease X, but when, in what order, and the time gaps between events. This allows it to capture patterns of disease progression over time.

The model can output forecasts of risk for more than 1,000 diseases, also giving estimates of when diseases are likely to occur (years or decades ahead).

Strengths and Advantages

Delphi-2M has several features that distinguish it:

1. Breadth of disease coverage: Unlike most current risk calculators that focus on one disease or a few (e.g. cardiovascular risk, cancer risk, diabetes risk), Delphi-2M forecasts many conditions—over 1,000.
2. Long forecasting horizon: It can project risks years, up to 10-20 years or more ahead, enabling early interventions.
3. Use of rich time-sequence data: Capturing the timing and order of medical events as well as intervals improves predictive power over models that just use static snapshots.
4. Training and validation across populations: The model was tested across two separate healthcare systems (UK and Denmark), which lends strength to its generalizability—though with caveats (see below).
5. Potential for preventive healthcare: By flagging high risks ahead of disease onset, the model offers opportunities for lifestyle modification, earlier screening, and possibly medical interventions—thus shifting focus from treating disease to preventing it.

Limitations, Risks, and Challenges

While promising, Delphi-2M (and tools like it) face several significant challenges.

1. Data bias and representativeness
   • The datasets used (UK Biobank, Danish registry) are large but may be biased in terms of ethnicity, socioeconomic factors, and health status. What works for cohorts in those populations may not transfer perfectly to more diverse or underrepresented groups.
   • The model may perform less well for diseases that are rare, have highly variable presentation, or for conditions less well recorded in medical records (e.g. mental health disorders, pregnancy‐related complications).
2. Accuracy and uncertainty over long timeframes
   • Predictions further into the future are intrinsically more uncertain. Although the model may do well for more predictable, slowly developing chronic conditions (heart disease, certain cancers), forecasting decades ahead has greater margin for error.
3. Interpretability
   • Generative AI models (even when modified) can be opaque. For both clinicians and patients, understanding why a certain risk is high—and what can be done about it—is crucial—but not always easy to extract from complex models.
4. Clinical readiness and regulatory issues
   • The model is not yet ready for routine clinical deployment. It must undergo further testing, including in more varied populations, and regulatory approvals.
   • There are privacy concerns, data security, and ethical issues in how such predictive models are used—e.g. risk of misuse by insurers, impact on mental health of patients told they are at high risk, etc.
5. Potential for overdiagnosis / unnecessary interventions
   • If a model predicts elevated risk for many diseases, there is a danger that patients may undergo many tests, possibly invasive ones, or begin treatments earlier than necessary—possibly with side‐effects—unless clear guidelines are established.
6. Variable performance across disease types
   • Model tends to do better with diseases with consistent progression patterns (e.g. cancer, diabetes, cardiac disease) than with those where progression is erratic, or heavily influenced by environmental shocks, or insufficiently documented.

Implications for Healthcare, Ethics, and Society

The arrival of tools like Delphi-2M has the potential to transform several dimensions of medicine and public health—but also raises significant ethical, logistical, and social questions.

Positive Impacts

• Preventive healthcare: Early risk detection allows interventions earlier—lifestyle changes, screenings, or prophylactic treatments—that may reduce disease burden.
• Personalized medicine: Each individual’s medical history, lifestyle, age etc., are used; risk predictions can be tailored, not one-size-fits‐all.
• Healthcare planning and resource allocation: Public health systems can use aggregate risk forecasts to anticipate needs (e.g. which regions/populations may have rising incidence of diabetes, cardiovascular disease) and direct preventive resources accordingly.
• Patient empowerment: Knowledge of risk may motivate healthy behaviors (quit smoking, change diet, increase exercise, etc.), if communicated responsibly.

Ethical, Legal, Social Risks

• Privacy and data protection: Use of large personal medical datasets must ensure anonymity, secure storage, consent, etc.
• Fairness and equity: Ensuring the model works well for all population groups, including those often underrepresented in medical data (ethnic minorities, lower-income populations, global south) is essential to avoid exacerbating health disparities.
• Psychological effects: Knowing one’s elevated risk mentally can cause anxiety, stress, or fatalism, especially if there are limited options for intervention.
• Insurance, employment, discrimination: Risk predictions might be misused by insurers or employers in ways that penalize individuals for predicted disease risk.
• Regulation and oversight: Who is responsible for errors, mispredictions? How are these tools approved for clinical use? What are liability issues?

Future Directions and Considerations

To realize the full promise of such models, certain steps will be important:

1. Broader validation – Testing in diverse populations, with different ethnic, socioeconomic, geographic backgrounds.
2. Integrating additional data types – Genetic/genomic data, proteomics, environmental exposures, social determinants, lifestyle in more detail could further improve predictions. Some work is already underway in related studies.
3. Model interpretability and explainability – Tools to show why a risk is elevated, what features contribute most, so that both clinician and patient understand and trust the prediction.
4. Clinical trials / intervention studies – To see whether acting on predictions (lifestyle modifications, screenings, treatments) actually reduces disease incidence or improves outcomes. Predicting risk is one thing; changing outcomes is another.
5. Ethical and regulatory frameworks – To govern how such tools are used, how data is handled, how risk information is shared, and how to avoid misuse.
6. Communication strategies – Clear, sensitive ways to communicate risk to patients, avoiding alarmism, supporting actionable advice.

Conclusion

Delphi-2M represents a milestone in the evolution of predictive medicine: a model that can use large, long-term, multidimensional health data to forecast risk for over a thousand diseases many years ahead. Its size, scope, and methodology suggest considerable promise for shifting medical care from reactive to preventive, for individual patients and for health systems.

However, its ultimate value will depend on careful validation, thoughtful integration into clinical workflows, protection of privacy and equity, and ensuring that predictions lead to positive, actionable changes rather than harm. As these challenges are addressed, tools like Delphi-2M could transform not only how we treat disease but how we think about health across a lifetime.