Better disease prediction, faster and at lower cost
- 70% Accuracy predicting disease 10 years ahead
- 2M Parameter model—small, fast, and cost-effective
- 3 days From publication to running live on Teradata
A 2-million-parameter model that could transform how healthcare systems plan
Researchers at the European Molecular Biology Laboratory and the German Cancer Research Center developed Delphi 2M—a small language model adapted to predict the diseases a patient will develop up to 10 years in the future, with up to 70% accuracy. Teradata's forward-deployed engineering team, leveraging nPath and Bring Your Own Model (BYOM) technologies, had the model running on Teradata just three days after it was published—demonstrating the speed and power of Teradata's AI platform for healthcare innovation.
- Delphi 2M repurposes transformer architecture to process longitudinal patient data
- The model predicts future disease onset across patient cohorts at aggregate scale
- Teradata's BYOM enables in-database deployment of task-specific small language models
- nPath technology enables powerful analysis of patient journey sequence data
Healthcare systems struggle to forecast demand as populations age and costs rise
Caring for aging populations across many geographies places mounting pressure on healthcare systems to plan services more accurately and cost-effectively. Traditional demand forecasting methods lack the predictive depth needed to anticipate which services different patient cohorts will require—and when. Meanwhile, deploying AI in large enterprises has historically been bottlenecked by the high cost of inference at scale, making large language models impractical for many real-world healthcare applications.
- Aging population demographics are straining healthcare planning and resource allocation
- Conventional forecasting cannot predict disease onset at the individual or cohort level
- Large language model inference costs are prohibitive for enterprise-scale healthcare AI
- Speed-to-deployment for emerging research models is typically slow and resource-intensive
BYOM and nPath bring Delphi 2M to life on Teradata in 3 days
Delphi 2M's researchers made a clever architectural innovation: replacing the positional encoding in a standard transformer model—which tracks word order in a sentence—with age encoding that tracks a patient's timeline. Fed with restructured longitudinal patient data, the resulting model can predict future disease onset with remarkable accuracy. Teradata's BYOM technology enabled the engineering team to deploy this small, fine-tuned model directly in-database, while nPath provided the sequencing analytics needed to structure patient journey data—all operational within three days of the paper's release.
- Positional encoding replaced with age encoding to process patient timelines as sequences
- Longitudinal patient data restructured and fed to the model for cohort-level predictions
- BYOM deploys the 2M-parameter model in-database—faster and cheaper to score than large models
- nPath handles complex patient sequence analytics natively within Teradata
AI-driven healthcare planning that is faster, cheaper, and already accurate enough to act on
While Delphi 2M is still early-stage research and not yet cleared for individual clinical use, it’s already accurate enough for aggregate demand forecasting—helping healthcare systems anticipate which services different cohorts will need across different regions. Running on Teradata, the model delivers this at a fraction of the cost and complexity of large language model alternatives, with the potential to fundamentally transform how health systems plan and fund care.
- 3 days from research publication to live deployment on Teradata—a record-speed proof of platform agility
- 70% accuracy in disease onset prediction 10 years ahead—sufficient for population-level healthcare demand planning
- Lower cost small model inference in database delivers enterprise AI at a fraction of large model running costs
接続しましょう
Teradata が、ビジネスの成果を加速し、必要なビジネスの俊敏性を提供するのにどのように役立つかをご覧ください。
当社の営業担当者がお手伝いします。