Integrative metabolic and inflammatory phenotyping identifies cancer cachexia independent of weight history

Cancer cachexia (CCx) is a complex, multifactorial syndrome associated with poor outcomes, yet its clinical identification is often hindered by unreliable recall of historical body weight. We developed an interpretable machine learning (ML) framework to detect CCx by integrating routinely collected metabolic and inflammatory hematological features from the nationwide multicenter INSCOC cohort in China ( N = 16,858). Using a systematic, data-driven strategies, we identified a stable and biologically coherent set of features that collectively characterize a cachexia-associated phenotype. The final model demonstrates consistent predictive discrimination for CCx in both test cohort (AUC: 0.758, 95% CI: 0.742-0.775) and validation cohort (AUC: 0.748, 95% CI: 0.730-0.766). Importantly, this phenotype-based model generalizes under strict temporal validation and enables objective identification of CCx in real-world clinical settings without reliance on recalled body weight. Together, our findings support a conceptual shift from a purely weight-based syndrome of CCx toward an integrative metabolic-inflammatory phenotype.