Artificial intelligence-enabled updating and application of guideline-based algorithms for symptom management in adults with cancer: a concordance evaluation of cancer-related fatigue, depression, and pain
Introduction Cancer-related fatigue, depression, and pain are highly prevalent, often co-occur and impair physical function, social activities, and quality of life among adults with cancer. Despite the availability of evidence-based clinical guidelines that provide recommendations for assessment and management, these symptoms remain under-recognized and undertreated. Clinical decision support systems (CDSS) can operationalize guideline-based symptom management algorithms at the point of care. However, maintaining these systems as evidence and practice evolve is labor-intensive, limiting adoption and sustainability. To address this gap, we sought to evaluate an artificial intelligence (AI)-enabled approach to updating guideline-based algorithms and generating patient-specific recommendations for managing fatigue, depression, and pain among adults with cancer. Methods This study evaluated AI performance (Gemini 3.0 Pro) for two tasks: identifying and interpreting updates to cancer symptom management guidelines for fatigue, depression, and pain, and producing casespecific recommendations aligned with an existing rule-based CDSS. Guideline updates were assessed against expert-judged reference standards using an ordinal rubric for detection, completeness, accuracy, and safety. Vignette-based recommendations were compared using precision, recall, and F1-score at the level of discrete clinical recommendation elements. Results For guideline updating, AI performance met pre-specified acceptability thresholds for fatigue and pain, with high completeness and accuracy and no clinical safety risk. Performance for depression was variable, driven by omission of guideline-supported integrative therapies and non-specific pharmacotherapy guidance. In vignette analyses, AI-generated recommendations showed strong concordance with the rule-based CDSS (F1 0.86–0.96). Most mismatches between AI and the CDSS came from AI leaving something out, not from giving unsafe advice. Discussion This suggests AI currently works best as a support tool rather than making clinical decisions on its own. Broader clinical use will require further work to test a greater number of test vignettes and its performance in real-world, multi-site settings with diverse patient populations.
Authors
- Kristin Röper
- Rachel A. Pozzar (ORCID: https://orcid.org/0000-0002-5033-5701)
- David F. Lobach (ORCID: https://orcid.org/0000-0001-6126-6441)
- Janet Lee Abrahm (ORCID: https://orcid.org/0000-0002-7737-1323)
- Marilyn J. Hammer (ORCID: https://orcid.org/0000-0002-9561-6144)
- Evan Michael Graboyes (ORCID: https://orcid.org/0000-0003-3766-468X)
- Mary Elizabeth Cooley (ORCID: https://orcid.org/0000-0001-5353-9542)
- James Shalaby
Institutions
- Brigham and Women's Hospital (US)
- Northeastern University (US)
- Medical University of South Carolina (US)
- Dana-Farber Cancer Institute (US)
Publication Details
- Journal
- Frontiers in Medicine
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3389/fmed.2026.1975675
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- article
- Field-Weighted Citation Impact
- 0.00