TrajOnco: A Multi-Agent Framework for Temporal Reasoning Over Longitudinal EHR for Multi-Cancer Early Detection
In the realm of healthcare, the ability to accurately estimate cancer risk from longitudinal electronic health records (EHRs) has the potential to revolutionize early detection and patient care. However, modeling the complex trajectories of patients remains a daunting challenge. To address this issue, researchers have developed TrajOnco, a training-free, multi-agent large language model (LLM) framework designed to facilitate scalable multi-cancer early detection.
Overview of TrajOnco
TrajOnco employs a sophisticated chain-of-agents architecture combined with long-term memory capabilities. This innovative design allows it to perform temporal reasoning over sequential clinical events, resulting in the generation of detailed patient-level summaries, evidence-linked rationales, and predicted risk scores. The framework represents a significant advancement in the field of cancer detection by leveraging the wealth of information contained within EHRs.
Evaluation and Performance
The effectiveness of TrajOnco was rigorously evaluated using de-identified Truveta EHR data across 15 different cancer types. The assessment involved matched case-control cohorts aimed at predicting the risk of cancer diagnosis within a year. In a zero-shot evaluation, TrajOnco achieved Area Under the Receiver Operating Characteristic (AUROC) scores ranging from 0.64 to 0.80. Remarkably, the framework performed comparably to traditional supervised machine learning approaches in a lung cancer benchmark, while also showcasing superior temporal reasoning capabilities compared to single-agent LLMs.
Advantages of Multi-Agent Design
The multi-agent design of TrajOnco not only enhances its performance but also allows for effective temporal reasoning using smaller-capacity models, such as GPT-4.1-mini. This flexibility broadens the accessibility of advanced cancer detection methodologies, making them applicable in various clinical settings without the necessity for extensive computational resources.
Validation and Interpretability
Human evaluation was conducted to validate the fidelity of TrajOnco’s output, ensuring that the generated predictions and rationales are both accurate and clinically relevant. Furthermore, TrajOnco’s interpretable reasoning outputs can be aggregated to uncover population-level risk patterns that align with established clinical knowledge, providing valuable insights for healthcare professionals.
Conclusion
The development of TrajOnco underscores the potential of multi-agent large language models to perform interpretable temporal reasoning over longitudinal EHRs. The framework not only advances scalable multi-cancer early detection but also enhances clinical insight generation, paving the way for improved patient outcomes in cancer care.
- Framework Name: TrajOnco
- Architecture: Multi-agent, long-term memory
- Evaluation Data: De-identified Truveta EHR data across 15 cancer types
- AUROC Scores: 0.64 – 0.80
- Comparison: Comparable to supervised machine learning, better temporal reasoning than single-agent LLMs