Why the study?
Most healthcare providers are unfamiliar with diagnostic criteria for cancer cachexia, leaving patients often undiagnosed.
Does a guideline-driven Clinical Reasoning AI model improve the identification of cancer cachexia compared to standard ICD coding, NLP, and vanilla LLMs?
Cross-Sectional (n=50)
Does a guideline-driven Clinical Reasoning AI model improve the identification of cancer cachexia compared to standard ICD coding, NLP, and vanilla LLMs?
A guideline-driven Clinical Reasoning AI model accurately identifies cancer cachexia from complex EHR data, outperforming standard ICD coding, keyword NLP, and vanilla LLMs.
Answers are grounded in this paper and cite their sources.
Enables scalable cachexia detection where provider familiarity is low; extends guideline criteria to AI-driven chart review in oncology.
1613 Background: Cancer cachexia, a severe wasting syndrome, affects up to 87% of cancer patients 1 and contributes to ≥20% of cancer deaths 2 . Despite this burden, ~85% of healthcare providers are unfamiliar with diagnostic criteria 3 , leaving patients often undiagnosed. This study evaluated whether a Clinical Reasoning AI model configured on four cachexia guidelines (Fearon 4 , Evans 5 , GLIM 6 and mGPS 7 criteria) and designed to emulate clinician chart review, could accurately identify cancer cachexia. Methods: Using structured (vitals, labs) and unstructured (clinical notes) electronic health record data the Clinical Reasoning AI employed retrieval-augmented generation 8 and agentic techniques with large language models (LLMs) to: (1) retrieve clinically relevant information based on the clinical guidelines via natural language processing (NLP) and embedding, and (2) apply guideline-based reasoning to classify patients into Group 1 (no cachexia features), Group 2 (meets ≥1 guideline regardless of documented diagnosis), or Group 3 (documented cancer cachexia). Accuracy was assessed against a gold standard of 50 patients classified by two independent clinicians, with conflicts resolved by a third. Performance was compared against baseline methods: ICD coding (ICD-9 799.4; ICD-10 C80.9), NLP using keywords like cancer or cachexia, and LLMs (Vanilla OpenAI GPT-o4-mini) prompted without explicit instruction to use guidelines. Results: The Clinical Reasoning AI demonstrated high precision, sensitivity, and specificity, outperforming baselines, particularly for Group 2 patients. ICD and NLP methods could not detect Group 2, due to code- and keyword-based constraints. Performance of LLMs without guidelines worsened with structured and unstructured data, indicating sensitivity to data overload and inability to prioritise clinically relevant information. Conclusions: These results show that embedding clinical guidelines within a structured retrieval-and-reasoning architecture enabled accurate identification of cancer cachexia, addressing a critical gap where standard coding and keyword methods fail. The guideline-driven approach maintained accurate clinical reasoning while processing complex EHR data. The Clinical Reasoning AI model is currently being deployed at the point of care to assess real-world implementation and impact on patient outcomes. Group 2 Group 3 AI Model Precision Sensitivity F1 Specificity Precision Sensitivity F1 Specificity ICD N/A N/A N/A N/A 0.759 0.815 0.786 0.696 Simple NLP N/A N/A N/A N/A 0.727 0.593 0.653 0.739 Vanilla OpenAI GPT-o4-mini (unstructured data only) 0.500 0.500 0.500 0.765 0.594 0.704 0.644 0.438 Vanilla OpenAI GPT-o4-mini (structured and unstructured data) 0.500 0.062 0.111 0.971 0.385 0.185 0.250 0.652 Clinical Reasoning AI with GPT-o4-mini 0.762 1.00 0.865 0.853 0.964 1.00 0.982 0.957
No takes yet. Share an insight, caveat, or question.
Gupta et al. (2026) conducted a cross-sectional in Cancer cachexia (n=50). Clinical Reasoning AI model vs. ICD coding, simple NLP, and vanilla LLMs was evaluated on Accuracy of identifying cancer cachexia (precision, sensitivity, F1, specificity). A Clinical Reasoning AI model accurately identified documented cancer cachexia with 1.00 sensitivity and 0.982 F1 score, outperforming standard ICD coding and simple NLP methods.
Synapse has enriched 10 related papers. Consider them for comparative context: