AI Model Accurately Predicts Multiple Disease Risks

On September 22, 2025, an innovative AI model known as Delphi-2M was reported, showcasing its capability to assess multiple disease risks using extensive medical history for predictions. This groundbreaking model has demonstrated that it can perform comparably to existing single-disease risk assessment tools, prompting further investigations into its practical applications in healthcare.

Introduction to Delphi-2M

Delphi-2M, developed by a group of European scientists, employs a model architecture based on GPT (Generative Pre-trained Transformer). Unlike other models designed to predict text, Delphi-2M is focused on forecasting health outcomes based on a patient's previous medical data. It has been trained on an extensive dataset from the UK Biobank, encompassing health information from approximately half a million individuals.

The importance of this model lies in its potential to transform how healthcare professionals anticipate and prepare for future health events. According to the study published in Nature, the one-of-a-kind model can predict outcomes for over 1,000 diseases simultaneously.

Mechanism of Action

Building on the strengths of existing models, Delphi-2M assimilates detailed patient data, including:

  • Age
  • Sex
  • Body Mass Index (BMI)
  • Tobacco use
  • Prior health diagnoses

This data serves as a prompt, allowing the model to generate predictions regarding various health events, such as the probability of developing pneumonia or heart failure at certain intervals.

Key Features and Advantages

The results yielded by the model include probabilities for future health events, which can help general practitioners make informed health predictions. For instance, the output might indicate:

  • Pneumonia: 18% risk
  • Heart Failure: 9% risk
  • Death: 3% risk

Additionally, the model not only aids in individual assessments but also facilitates population modeling. By simulating health trajectories based on demographic data, it offers insights into potential healthcare burdens such as hospitalizations, deaths, and the incidence of diseases.

Comparison with Existing Models

Delphi-2M has shown a remarkable ability to generally surpass a multi-disease predictor that was based on 67 biomarkers identified from UK Biobank data. Nevertheless, it is worth noting that while the model's disease predictions were impressive, they still fell short when compared with specific biomarkers like HbA1c, which is commonly used for diabetes risk assessment.

Limitations and Challenges

The creators of Delphi-2M recognized various limitations within their framework:

  1. Data Age: Since the UK Biobank primarily recruits individuals aged 40-70, data for individuals over 80, who may have differing health trajectories, is severely limited.
  2. Data Absence Bias: Health histories missing from the dataset due to the mortality of participants could skew the predictive accuracy.
  3. Data Scarcity: A general shortage of comprehensive health data across various demographics poses challenges for training more robust models.

Innovative Solutions to Data Limitations

To address the challenges of data scarcity, researchers have explored alternatives such as utilizing synthetic data generated by the model itself:

  • Synthetic Patient Histories: By using the Delphi-2M model to create simulated data, researchers have trained new models with almost equivalent accuracy compared to those trained on real data.

This innovative approach could mitigate privacy concerns associated with using real patient data while still leveraging the predictive capabilities of AI.

Conclusion

While the Delphi-2M model cannot predict future health events with absolute certainty, it represents a significant advancement in the use of AI for health predictions. By creating risk stratification and developing projections based on various health factors, this model paves the way for more personalized healthcare strategies.

Literature Cited

[1] Shmatko, A., Jung, A. W., Gaurav, K., Brunak, S., Mortensen, L. H., Birney, E., … & Gerstung, M. (2025). Learning the natural history of human disease with generative transformers. Nature, 1-9.

If you wish to support our independent journalism, consider making a donation. Every contribution, no matter how small, helps sustain our reporting and educational efforts.