Back to AI Research

AI Research

SAGEAgent: A Self-Evolving Agent for Cost-Aware Mod... | AI Research

Key Takeaways

  • SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction In clinical oncology, patients often undergo a series...
  • Does every cancer patient truly need a complete diagnostic workup for accurate survival prediction?
  • In multimodal clinical oncology, diagnostic modalities follow a clinically mandated order of escalating burden -- from demographics collected at intake to genomic profiling requiring specialized tissue analysis.
  • SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction
  • In clinical oncology, patients often undergo a series of diagnostic tests—such as demographics, imaging, pathology, and genomic profiling—that increase in cost and physical burden.
Paper AbstractExpand

Does every cancer patient truly need a complete diagnostic workup for accurate survival prediction? In multimodal clinical oncology, diagnostic modalities follow a clinically mandated order of escalating burden -- from demographics collected at intake to genomic profiling requiring specialized tissue analysis. Current multimodal survival methods either assume all modalities are available or passively handle missing data, but none actively reason about whether acquiring the next modality is justified for a given patient along this ordered workflow. We formulate this as a sequential decision problem and propose SAGEAgent (Sequential Acquisition Guided by Experience), a self-evolving LLM-based clinical agent that decides which diagnostic modalities to acquire for each patient, balancing predictive accuracy against clinical invasiveness. SAGEAgent reasons about each patient's evolving diagnostic state through clinical tools that translate numerical predictions into text, an episodic memory that retrieves similar past cases, and a semantic memory that accumulates reusable decision patterns from experience. Experiments on a glioma cohort combining TCGA-LGG, TCGA-GBM, and BraTS with four diagnostic modalities demonstrate that SAGEAgent achieves competitive survival prediction accuracy while reducing average acquisition burden by 55%.

SAGEAgent: A Self-Evolving Agent for Cost-Aware Modality Acquisition in Multimodal Survival Prediction
In clinical oncology, patients often undergo a series of diagnostic tests—such as demographics, imaging, pathology, and genomic profiling—that increase in cost and physical burden. Current medical AI models typically require all these tests to be completed to provide a survival prediction, or they struggle to handle missing data effectively. SAGEAgent is a new AI framework designed to act as a clinical assistant that decides, step-by-step, whether a patient truly needs the next diagnostic test or if a reliable prediction can be made with the information already gathered. By balancing the need for accuracy against the burden of invasive procedures, the agent aims to streamline the diagnostic process.

How SAGEAgent Makes Decisions

SAGEAgent operates like a human clinician following a logical workflow. It uses a frozen Large Language Model (LLM) that does not require further training to function. At each stage of the diagnostic process, the agent evaluates three key sources of information:

  • Clinical Tools: These translate complex numerical data, such as risk scores and uncertainty levels, into plain language that the LLM can interpret.

  • Episodic Memory: The agent retrieves data from past patients who had similar clinical profiles, allowing it to see how those cases were handled and what the outcomes were.

  • Semantic Memory: This component stores reusable decision rules that the agent has learned over time, helping it apply consistent logic to new patients.

The Self-Evolution Mechanism

A unique feature of SAGEAgent is its ability to "self-evolve" without needing traditional gradient-based updates. The agent periodically reflects on its own performance by reviewing past decisions. It identifies which actions led to successful outcomes and which did not, then updates its semantic memory with new, validated rules. This creates a closed-loop system where the agent continuously refines its decision-making process based on experience, effectively learning which diagnostic steps are most valuable for different types of patients.

Results and Clinical Impact

When tested on a large cohort of glioma patients, SAGEAgent demonstrated that it could maintain high levels of survival prediction accuracy while significantly reducing the diagnostic burden. Specifically, the agent was able to reduce the average number of required tests by 55% compared to a full diagnostic workup. Furthermore, the agent provides a "chain-of-thought" explanation for every decision, ensuring that clinicians can audit the reasoning behind why a specific test was ordered or skipped. This transparency is intended to build trust in AI-assisted clinical workflows.

Key Takeaways

The research suggests that not every patient requires a complete diagnostic workup to receive an accurate prognosis. By using a combination of past case experiences and learned clinical rules, SAGEAgent can identify when a patient has enough diagnostic information to proceed to treatment planning. While the current study focused on glioma, the framework’s ability to respect clinical ordering and provide auditable reasoning offers a promising path for applying cost-aware AI to other areas of oncology and medicine.

Comments (0)

No comments yet

Be the first to share your thoughts!