HemOncAgent: an artificial intelligence system for retrieving structured and narrative oncology knowledge

In the complex world of cancer treatment, doctors must navigate two very different kinds of information. One type is rigid and precise, like a catalog of ingredients that lists exactly which drugs belong in a specific combination or which medical codes describe a particular disease. The other type is fluid and descriptive, found in the written notes of specialists who explain how to sequence treatments over time, deciding when to start a new therapy or how to manage a patient after a recurrence. For decades, computer systems designed to help doctors have struggled to handle both at once. Some excel at finding exact facts but fail to understand the story of a treatment plan, while others are good at reading text but miss the critical connections between specific medicines. This gap matters because a doctor asking a question about cancer care might need a precise drug list one moment and a narrative explanation of treatment pathways the next.

A team of researchers has built a new digital assistant called HemOncAgent to bridge this divide. Instead of forcing a single method to do all the work, this system acts like a thoughtful librarian who knows when to check a structured database and when to read a detailed handbook. The researchers tested this tool against standard computer systems using hundreds of questions about oncology knowledge. They found that by letting the computer choose the right source for each specific question, the new system could answer both factual queries and complex treatment scenarios with high accuracy. The results suggest that for medical knowledge that is part hard data and part human experience, a hybrid approach is far more reliable than relying on just one type of search.

The challenge in oncology is that knowledge exists in two distinct forms. On one hand, there are curated databases that store structured facts, such as which drugs make up a specific regimen or what standard medical codes apply to a condition. These are like the hard rules of a game, where every element has a defined place. On the other hand, there are resources written by clinicians that describe how treatments are actually used in practice, detailing the order in which therapies are given for different stages of a disease. These narratives capture the nuance of decision-making that rigid databases often miss. While large language models—computer programs trained on vast amounts of text—can answer questions, they often hallucinate or guess when they lack access to these specific external sources. To solve this, researchers developed a system that can reach out to both types of resources, selecting the most appropriate one based on what the user is asking.

The researchers created HemOncAgent to act as a central hub that connects to two specific knowledge sources. The first is a structured knowledge graph, a massive network of over 126,000 concepts and 424,000 relationships linking drugs, regimens, and diseases. This source is excellent for answering questions about exact compositions or regulatory approvals. The second source is a narrative website containing thousands of pages of text written by experts, which describes treatment pathways and care sequences. When a question is asked, the agent analyzes the request and decides whether to query the structured graph, search the narrative text, or use both. It then retrieves the relevant information and synthesizes a final answer. To see if this approach worked, the team compared HemOncAgent against three other systems: a standard language model with no outside knowledge, a system that only searched the narrative text, and a system that only queried the structured graph.

The testing involved 600 questions divided into two groups. The first group focused on structured facts, such as identifying the drugs in a regimen or finding a specific medical code. The second group focused on narrative details, such as determining the order of treatments for a specific disease setting. The results showed a clear pattern. The system that only searched the narrative text performed well on treatment sequences but failed miserably at finding exact codes or drug lists. Conversely, the system that only used the structured graph was excellent at facts but struggled to find the right treatment pathways described in text. The standard language model without any external help performed poorly on both. HemOncAgent, however, achieved the highest overall score. It matched the performance of the specialized graph system on factual questions and outperformed the text-only system on narrative questions, successfully combining the strengths of both approaches.

When the researchers looked closer at the specific types of questions, the advantages of the hybrid system became even clearer. For questions requiring exact matches, such as finding a drug's classification or a clinical trial number, the agent relied almost entirely on the structured database, achieving near-perfect accuracy. For questions about treatment settings, such as what happens after a first-line therapy fails, the agent primarily searched the narrative pages. In these cases, it significantly outperformed the graph-only system, which could not find the answers hidden in the prose. The agent also demonstrated that it could intelligently mix sources when necessary, using both the graph and the text for about a quarter of the narrative questions to ensure a complete answer. This flexibility allowed it to maintain high accuracy across the entire range of oncology knowledge.

The study also tested how robust these findings were by changing the underlying computer model that generated the answers. Even when the researchers swapped the primary engine for a different type of artificial intelligence, the hybrid system continued to outperform the single-source alternatives. While the absolute scores changed slightly, the pattern remained: having access to both structured and narrative sources provided a consistent advantage. The researchers noted that the automated grading system used to evaluate the answers agreed with a second, independent grading system on nearly 94 percent of the responses, giving confidence that the results were not just a fluke of one specific evaluation method.

Despite these successes, the authors are careful to note the boundaries of their findings. The questions used for testing were generated directly from the same resources the system was searching, meaning the test measured how well the system could retrieve information from those specific sources rather than how well it could reason about new, unseen clinical scenarios. The study does not claim that this system is ready to replace doctors or that it can answer every possible question about cancer care, especially since medical guidelines change rapidly and the knowledge base represents a snapshot in time. However, the results strongly suggest that for clinical knowledge resources that span both hard data and human narrative, a single search strategy is insufficient. By allowing an artificial intelligence agent to choose the right tool for the job, the researchers have demonstrated a practical way to make medical information more accessible and reliable.