Machine Learning Highlights Association of Body Composition with Cholangiocarcinoma Mortality

Background Cholangiocarcinoma (CCA) is a high mortality malignancy. Previous studies have largely overlooked individual differences in physiological and metabolic characteristic, limiting the ability to explain and predict the high mortality of CCA. Body composition (BC), a key manifestation of these characteristics, plays a crucial role in cancer. However, its impact on CCA mortality remains inadequately understood.

Methods All data were obtained from the UK Biobank, with CCA diagnosed based on ICD-10 codes. Proportional hazards models were used to assess potential associations between BC and CCA and its subtypes, intrahepatic (ICC) and extrahepatic (ECC). Additionally, multiple machine learning algorithms were employed to develop prognostic prediction models for CCA. And the SHapley Additive exPlanations approach was applied to improve model interpretability.

Results Fat mass in the arm (left: HR=1.31, right: HR=1.28, P<0.05) and leg (left: HR=1.31, right: HR=1.28, P<0.05) were positively associated with total CCA and ICC mortality. In contrast, ECC mortality was associated with leg fat-free mass (left: HR=2.21, right: HR=2.10, P<0.05). Additionally, the eXtreme Gradient Boosting model, which centered on BC, outperformed all other models, achieving area under the curve values of 0.884 for total CCA, 0.865 for ICC, and 0.737 for ECC in the test set.

Conclusion This study highlights significant associations between site-specific BC and CCA mortality and develops a CCA mortality prediction model centered on BC variables. These findings provide valuable insights for personalized clinical management and prognostic interventions.

Publication History

Article published online:
20 January 2025