Structured reasoning failures compromise LLM interpretation of clinical oncology notes

Area of Science:

• Artificial intelligence in medicine
• Oncology clinical decision support

Background:

• Large language models (LLMs) demonstrate high performance on clinical benchmarks.
• The reliability of LLM reasoning in real-world oncology settings is not well understood.

Purpose of the Study:

• To evaluate the reasoning reliability of LLMs in authentic oncology clinical notes.
• To identify and categorize reasoning errors, including cognitive biases, in LLM interpretations.

Main Methods:

• Utilized a novel hierarchical error taxonomy to analyze LLM reasoning on retrospective oncology notes (breast, pancreatic, prostate cancer).
• Compared error rates and patterns between GPT-4 and GPT-5.1.
• Assessed the effectiveness of automated LLM evaluators and a self-mitigation strategy.

Main Results:

• GPT-4 exhibited reasoning errors in 23.1% of interpretations, predominantly linked to cognitive biases like confirmation and anchoring bias.
• Errors were more common in recommendation tasks and associated with guideline-discordant recommendations and lower clinical impact.
• GPT-5.1 showed reduced errors but still displayed structured reasoning failures.
• Automated evaluators detected errors but struggled with subtype classification; self-mitigation offered limited improvement.

Conclusions:

• Endpoint accuracy may obscure clinically significant LLM reasoning failures.
• Cognitive biases and omission errors are strongly linked to potentially harmful outputs in oncology LLM applications.
• Rigorous evaluation and continuous monitoring of LLM reasoning fidelity are essential prerequisites for safe implementation in oncology decision support.