Journal of cachexia, sarcopenia and muscle · 2026
Early Identification and Prognostic Stratification of Cancer Cachexia Using Explainable Machine Learning: A Multicentre Cohort Study.
СтатьяМногоцентровое исследованиеНаблюдательное исследованиеНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Huang Wei, Pan Kan, Zhao Mingjian, Zhao Tingting, Xu Nuo, Shi Hanping
Аннотация
BACKGROUND: Cancer cachexia is a heterogeneous syndrome that remains frequently underrecognized in routine oncology care, especially before overt wasting develops. We aimed to develop and externally validate an explainable two-stage machine learning framework for early cachexia identification and prognostic stratification using routinely available clinical data.
METHODS: In this multicentre real-world study based on the INSCOC registry, 10 796 hospitalized adults with cancer were analysed, including 3995 patients with cachexia and 6801 without cachexia. First-stage machine learning models were trained separately in high- and low-prevalence settings to identify cachexia. Second-stage survival models were developed in 1711 cachexia patients with follow-up data from a 5172-patient follow-up cohort. A total of 586 deaths were observed in the cachexia follow-up cohort. External validation used an independent Fujian cohort of 886 patients.
RESULTS: Cachexia prevalence was 37.0% (3995/10 796). Compared with non-cachexia patients, cachexia patients had lower median BMI (21.0 vs. 23.4 kg/m2), albumin (37.9 vs. 39.7 g/L), triceps skinfold thickness (13.0 vs. 16.0 mm), calf circumference (33.0 vs. 35.0 cm) and higher CRP (5.9 vs. 3.9 mg/L); all p < 0.001. For cachexia identification, XGBoost achieved the best discrimination, with an AUC of 0.867 (95% CI: 0.857-0.878), sensitivity 0.706 and specificity 0.846 in the high-prevalence group, and an AUC of 0.870 (95% CI: 0.861-0.880), sensitivity 0.817 and specificity 0.756 in the low-prevalence group. In external validation, the AUCs were 0.744 in the high-prevalence subgroup and 0.694 in the low-prevalence subgroup. Among cachexia patients with follow-up, the 1-year mortality rate was 20.8% (95% CI: 19.0%-22.9%). For prognostic stratification, the random survival forest model showed the most favourable overall performance, with a C-index of 0.680 and time-dependent AUCs of 0.800, 0.777, 0.741 and 0.716 at 3, 6, 9 and 12 months, respectively. Advanced TNM stage and higher CRP were associated with worse survival, whereas higher albumin, prealbumin and HDL were associated with better survival.
CONCLUSIONS: This explainable two-stage framework enables accurate cachexia identification across different prevalence settings and clinically meaningful prognostic stratification in patients with established cachexia. Using routine variables and independent external validation, it provides a practical basis for earlier recognition, risk-adapted supportive care and precision management in oncology practice.
METHODS: In this multicentre real-world study based on the INSCOC registry, 10 796 hospitalized adults with cancer were analysed, including 3995 patients with cachexia and 6801 without cachexia. First-stage machine learning models were trained separately in high- and low-prevalence settings to identify cachexia. Second-stage survival models were developed in 1711 cachexia patients with follow-up data from a 5172-patient follow-up cohort. A total of 586 deaths were observed in the cachexia follow-up cohort. External validation used an independent Fujian cohort of 886 patients.
RESULTS: Cachexia prevalence was 37.0% (3995/10 796). Compared with non-cachexia patients, cachexia patients had lower median BMI (21.0 vs. 23.4 kg/m2), albumin (37.9 vs. 39.7 g/L), triceps skinfold thickness (13.0 vs. 16.0 mm), calf circumference (33.0 vs. 35.0 cm) and higher CRP (5.9 vs. 3.9 mg/L); all p < 0.001. For cachexia identification, XGBoost achieved the best discrimination, with an AUC of 0.867 (95% CI: 0.857-0.878), sensitivity 0.706 and specificity 0.846 in the high-prevalence group, and an AUC of 0.870 (95% CI: 0.861-0.880), sensitivity 0.817 and specificity 0.756 in the low-prevalence group. In external validation, the AUCs were 0.744 in the high-prevalence subgroup and 0.694 in the low-prevalence subgroup. Among cachexia patients with follow-up, the 1-year mortality rate was 20.8% (95% CI: 19.0%-22.9%). For prognostic stratification, the random survival forest model showed the most favourable overall performance, with a C-index of 0.680 and time-dependent AUCs of 0.800, 0.777, 0.741 and 0.716 at 3, 6, 9 and 12 months, respectively. Advanced TNM stage and higher CRP were associated with worse survival, whereas higher albumin, prealbumin and HDL were associated with better survival.
CONCLUSIONS: This explainable two-stage framework enables accurate cachexia identification across different prevalence settings and clinically meaningful prognostic stratification in patients with established cachexia. Using routine variables and independent external validation, it provides a practical basis for earlier recognition, risk-adapted supportive care and precision management in oncology practice.