DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System
DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System
Authors: Yunying Zhu, Andrew R. Weckstein, Kueiyu Joshua Lin, Jie Yang Date: 2026-05-14 Paper ID: openalex:2605.14227
Summary
DT-Transformer is a foundation model designed for longitudinal disease trajectory prediction within large, multi-hospital health systems. By leveraging 57.1M structured EHR entries from the Mass General Brigham network, the model effectively addresses the limitations of single-hospital research cohorts. It achieves high discriminative performance across nearly 900 disease categories, providing a robust framework for real-world clinical forecasting and early intervention.
Key Contributions
- Introduces DT-Transformer, a foundation model for EHR data trained on a large-scale, multi-hospital dataset of 1.7M patients.
- Achieves a median AUC of 0.871 for next-event prediction across 896 clinical disease categories.
- Demonstrates the scalability and generalizability of health system-wide EHR foundation models in both retrospective and prospective validation settings.
Open Questions & Future Work
Archivist Review
The paper describes a large-scale application of Transformer architectures to EHR data. I approved two open questions that characterize specific, critical bottlenecks in clinical foundation modeling: the performance decay over long horizons and the transition from predictive AUC performance to actionable clinical utility. No new concepts were approved as the model architecture is a standard Transformer application to EHR sequences.
Approved Open Questions
- Long-horizon clinical forecasting decay: This addresses a core bottleneck in the deployment of foundation models for disease progression, where long-term clinical utility depends on maintaining predictive accuracy beyond the next immediate event.
- Clinical utility evaluation paradigm: Transitioning from predictive discriminative performance to actionable clinical utility is a fundamental challenge for the long-term viability of clinical foundation models.
Rejected Candidates
- [open_question] Improving Long-Horizon Forecasting (
improving-long-horizon-forecasting) - other: Renamed to better reflect the underlying research bottleneck and avoid generic phrasing. - [open_question] Validating Clinical Utility (
clinical-utility-validation) - other: Renamed to better reflect the methodological nature of the evaluation challenge.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.14227
- paper_id
- 2605.14227
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-17T06:08:44Z