Am J Clin Nutr. 2022 Sep 12. pii: nqac251. [Epub ahead of print]
Liangyu Yin,
Jiuwei Cui,
Xin Lin,
Na Li,
Yang Fan,
Ling Zhang,
Jie Liu,
Feifei Chong,
Chang Wang,
Tingting Liang,
Xiangliang Liu,
Li Deng,
Mei Yang,
Jiami Yu,
Xiaojie Wang,
Minghua Cong,
Zengning Li,
Min Weng,
Qinghua Yao,
Pingping Jia,
Zengqing Guo,
Wei Li,
Chunhua Song,
Hanping Shi,
Hongxia Xu.
BACKGROUND: Diagnosing cancer cachexia relies extensively on the patient-reported historic weight, and failure to accurately recall this information can lead to severe underestimation of cancer cachexia.
OBJECTIVES: The present study aimed to develop inexpensive tools to facilitate the identification of cancer cachexia in patients without weight loss information.
METHODS: This multicenter cohort study included 12774 patients with cancer. Cachexia was retrospectively diagnosed using Fearon's framework. Baseline clinical features, excluding weight loss, were modeled to mimic a situation where the patient is unable to recall their weight history. Multiple machine learning (ML) models were trained using 75% of the study cohort to predict cancer cachexia, with the remaining 25% of the cohort used to assess model performance.
RESULTS: The study enrolled 6730 males and 6044 females (median age = 57.5 years). Cachexia was diagnosed in 5261 (41.2%) patients and most diagnoses were made based on the weight loss criterion. A 15-variable logistic regression (LR) model mainly comprising cancer types, gastrointestinal symptoms, tumor stage and serum biochemistry indices was selected among the various ML models. The LR model showed good performance for predicting cachexia in the validation data (area under the curve = 0.763, 95% confidence interval=[0.747, 0.780]). The calibration curve of the model demonstrated good agreement between predictions and actual observations (accuracy = 0.714, Kappa = 0.396, sensitivity = 0.580, specificity = 0.808, positive predictive value = 0.679, negative predictive value = 0.733). Subgroup analyses showed that the model was feasible in patients with different cancer types. The model was deployed as an online calculator and a nomogram, and was exported as predictive model markup language to permit flexible, individualized risk calculation.
CONCLUSIONS: We developed a ML model that can facilitate the identification of cancer cachexia in patients without weight loss information, which might improve decision-making and lead to the development of novel management strategies in cancer care.
Keywords: cancer cachexia; gastrointestinal symptoms; machine learning; weight loss