Abstract

Abstract Cachexia is a debilitating syndrome characterized by loss of skeletal muscle tissue that affects over 50% of cancer patients. Defining cachexia for the purposes of epidemiologic analyses has historically been determined by specific weight loss over specific time, and more recently relied on technically challenging interpretation of imaging studies. Subjective appetite or weight loss (AWL) as documented from clinician notes may offer another method of defining a similar phenotype. Although annotation of free-text notes at scale is challenging, we hypothesized that natural language processing (NLP) might successfully annotate AWL from cancer diagnosis medical notes and that such annotations would follow patterns expected from classic cachexia studies. We created a gold-standard dataset for NLP training and validation by manually labeling AWL in a cohort of 762 free-text initial consultation notes that were extracted from lung and pancreatic adenocarcinoma electronic health records. Using these labels, we fine-tuned a pre-trained neural Longformer model to classify whether a patient has symptoms of cachexia. The AWL model was evaluated on a cohort of 391 pan-cancer notes independent of the training cohort and achieved an AUC of 0.92, precision of 0.88, recall of 0.95, and accuracy of 0.94. To test whether the labels generated by this model behaved as expected from classic cachexia studies, we applied the model to 46,980 initial consultation notes from a pan-cancer cohort of patients with tumor genomic profiling from MSK-IMPACT, an FDA-authorized tumor sequencing assay. Overall, AWL was present in 28% of the patients and it was associated with gastrointestinal cancers, which is expected from previous studies of cachexia. Our analysis revealed that esophagogastric and pancreatic cancer patients had the highest rates of cachexia at around 67%. We also observed that patients with cachexia symptoms have a reduced overall survival compared to patients without (HR=1.80, 95% CI=[1.75, 1.85]), which held in several specific cancers including NSCLC (HR=1.79, 95% CI=[1.66, 1.94]), pancreatic cancer (HR=1.34, 95% CI=[1.22, 1.47]), and colorectal cancer (HR=1.71, 95% CI=[1.55, 1.89]). Cachexia symptoms were more prevalent in male patients than in female patients. Underweight patients by BMI at presentation were most likely to have cachexia symptoms. These results reciprocate well-established information about cancer cachexia, demonstrating that our text classification NLP model can be reliably used to predict AWL from free-text medical notes. AWL may thus be a viable means of studying correlates of cachexia at scale in real-world cohorts. Citation Format: Tricia Park, Karl Pichotta, Christopher J. Fong, Nikolaus Schultz, Puneeth Iyengar, Justin Jee, Ed Reznik. Automatic identification of subjective appetite or weight loss in clinician notes empowers studies of cachexia [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 911.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call