Truace
Subscribe

TrajOnco: a multi-agent framework for temporal reasoning over longitudinal EHR for multi-cancer early detection

Abstract Accurate estimation of cancer risk from longitudinal electronic health records (EHRs) could support earlier detection and improved care, but modeling such complex patient trajectories remains challenging. We present TrajOnco, a training-free, multi-agent large language model (LLM) framework designed for scalable multi-cancer early detection. Using a chain-of-agents architecture with long-term memory, TrajOnco performs temporal reasoning over sequential clinical events to generate patient-level summaries…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1596· cite & verify
TrajOnco: a multi-agent framework for temporal reasoning over longitudinal EHR for multi-cancer early detection
Emoji Use in the Electronic Health Record by David A. Hanauer, Gavin C. Raab, Shira N. Hanauer, Lisa Ferguson, Kellen McClain, Guan Wang, Michelle Rozwadowski, Sung W. Choi. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

In brief

Researchers built TrajOnco, a chain-of-agents large language model framework that reads sequences of clinical events in electronic health records and produces patient-level summaries, rationales linked to evidence, and risk scores. It was tested on de-identified Truveta data using matched case-control cohorts for 15 cancer types, predicting diagnosis within one year.

The work suggests that LLM agents could support scalable multi-cancer early detection without task-specific training, while also offering interpretable reasoning that clinicians can inspect. The abstract describes retrospective evaluation on matched cohorts, so real-world screening benefit has not been shown, and the lower end of the AUROC range indicates uneven performance across cancers.

Main points

  1. The framework is training-free and uses multiple LLM agents with long-term memory to reason over patient event timelines.
  2. In zero-shot tests across 15 cancer types, AUROCs ranged from 0.64 to 0.80.
  3. It performed comparably to supervised machine learning in a lung cancer benchmark and showed better temporal reasoning than single-agent LLMs.
  4. Smaller models such as GPT-4.1-mini could still carry out effective temporal reasoning in the multi-agent design.
  5. Human evaluation supported the groundedness of outputs, and aggregated rationales matched established clinical knowledge on population-level risk patterns.

The gain

A training-free multi-agent LLM system predicted 1-year cancer diagnosis risk from longitudinal health records, reaching AUROCs of 0.64–0.80 and matching supervised machine learning in a lung cancer benchmark.

The rundown

TrajOnco uses a chain-of-agents architecture in which agents process sequential clinical events and keep long-term memory. Each patient receives a summary, rationales tied to evidence in the record, and a predicted risk score for cancer diagnosis at one year.

The evaluation used de-identified Truveta EHR data and matched case-control cohorts covering 15 cancer types. Under zero-shot conditions, the system was compared with supervised machine learning on lung cancer and with single-agent LLMs on temporal reasoning.

The authors also report that the multi-agent design allowed smaller-capacity models to reason effectively, and that human reviewers found outputs grounded. Aggregating the reasoning outputs revealed population-level risk patterns consistent with established clinical knowledge, pointing to a use for clinical insight generation.

Sources

  1. Peer-reviewednpj Digital Medicine2026-10-15

The debate