Structured reasoning failures compromise LLM interpretation of clinical oncology notes
6 affiliations
- 1Mayo Clinic College of Medicine and Science
- 2School of Computing and AI, Arizona State University
- 3Department of Radiation Oncology, Mayo Clinic
- 4Mayo Clinic Comprehensive Cancer Center
- 5Department of Artificial Intelligence and Informatics, Mayo Clinic
- 6Dana-Farber Cancer Institute
Abstract
Large language models (LLMs) show strong performance on clinical benchmarks, yet their reasoning reliability in real-world oncology care remains unclear. We evaluated LLM reasoning on authentic oncology notes using a novel hierarchical error taxonomy across two retrospective cohorts spanning breast, pancreatic, and prostate cancer. GPT-4 produced reasoning errors in 23.1% of note interpretations, the majority reflecting cognitive bias patterns. Errors were more frequent in recommendation tasks and were strongly associated with guideline-discordant recommendations and lower clinician-rated clinical impact scores. Confirmation bias, anchoring bias, and omission errors were most strongly linked to potentially harmful outputs. Compared to GPT-4, GPT-5.1 demonstrated reduced error rates and improved clinical performance but retained structured reasoning failure patterns. Automated LLM-based evaluators detected error presence but failed to reliably classify subtypes, and a preliminary self-mitigation strategy yielded only modest improvement. Endpoint accuracy alone may mask clinically meaningful reasoning failures. Therefore, evaluating and monitoring reasoning fidelity should be a prerequisite for safe deployment of LLMs in oncology decision support.
Related
- Enhancing Evidence Translation Through Structured Clinical Reasoning: A Mixed-Methods Controlled Feasibility Study of Peer Visibility and Readiness Mechanisms in Applied Learning
- Implausible 100-400 m Performance Times and Inadequate Testing Distances Compromise the Estimation and Interpretation of Critical Velocity. Comment on Amara et al. Relationships Between Dry-Land Load-Velocity Parameters and In-Water Bioenergetic Performance in Competitive Swimmers. Sports 2026, 14, 11.
- Teacher-student-machine interaction autonomous learning: a structured LLM-integrated framework for developing independent clinical reasoning in residency training.
- Cognitive interpretation, behaviors, and localized somatic manifestations: can a structured link be conceptualized in light of P.N.E.I.?
- Automated Interpretation of EEG Reports Using a Large Language Model with Structured Confidence Outputs
BibTeX
@article{kenaston2026structured,
title = {{Structured reasoning failures compromise LLM interpretation of clinical oncology notes}},
author = {Matthew W. Kenaston and Umair Ayub and Mihir Parmar and Priya Kumar and Muhammad Umair Anjum and Syed Arsalan Ahmed Naqvi and Aadel A. Chaudhuri and Yousef Zakharia and Elisabeth I. Heath and Tanios S. Bekaii-Saab and Cui Tao and Eliezer M. Van Allen and Ben Zhou and YooJung Choi and Chitta Baral and Irbaz Bin Riaz},
journal = {npj Digital Medicine},
year = {2026},
month = jun,
doi = {10.1038/s41746-026-02951-5},
url = {https://doi.org/10.1038/s41746-026-02951-5},
note = {CC BY},
}
RIS
TY - JOUR
TI - Structured reasoning failures compromise LLM interpretation of clinical oncology notes
AU - Matthew W. Kenaston
AU - Umair Ayub
AU - Mihir Parmar
AU - Priya Kumar
AU - Muhammad Umair Anjum
AU - Syed Arsalan Ahmed Naqvi
AU - Aadel A. Chaudhuri
AU - Yousef Zakharia
AU - Elisabeth I. Heath
AU - Tanios S. Bekaii-Saab
AU - Cui Tao
AU - Eliezer M. Van Allen
AU - Ben Zhou
AU - YooJung Choi
AU - Chitta Baral
AU - Irbaz Bin Riaz
PY - 2026
DA - 2026/06/01
JO - npj Digital Medicine
DO - 10.1038/s41746-026-02951-5
UR - https://doi.org/10.1038/s41746-026-02951-5
AB - Large language models (LLMs) show strong performance on clinical benchmarks, yet their reasoning reliability in real-world oncology care remains unclear. We evaluated LLM reasoning on authentic oncology notes using a novel hierarchical error taxonomy across two retrospective cohorts spanning breast, pancreatic, and prostate cancer. GPT-4 produced reasoning errors in 23.1% of note interpretations, the majority reflecting cognitive bias patterns. Errors were more frequent in recommendation tasks and were strongly associated with guideline-discordant recommendations and lower clinician-rated clinical impact scores. Confirmation bias, anchoring bias, and omission errors were most strongly linked to potentially harmful outputs. Compared to GPT-4, GPT-5.1 demonstrated reduced error rates and improved clinical performance but retained structured reasoning failure patterns. Automated LLM-based evaluators detected error presence but failed to reliably classify subtypes, and a preliminary self-mitigation strategy yielded only modest improvement. Endpoint accuracy alone may mask clinically meaningful reasoning failures. Therefore, evaluating and monitoring reasoning fidelity should be a prerequisite for safe deployment of LLMs in oncology decision support.
SN - 2398-6352
ER -