Lumbar centroid level subcutaneous fat volume increased performance of prognostic predictive model in digestive...

Nutritional indicators play an important role in predicting the prognosis of digestive system cancers. Measures of adipose tissue distribution derived from computed tomography (CT), such as subcutaneous fat volume, are promising for assessing systemic inflammation and nutritional status. However, their integration into standardized prognostic models is still limited. This study aimed to increase the performance of the Cox regression model (Coxm) by adding the third lumbar vertebra centroid level subcutaneous fat volume (L3 CLSFV) and to assess its influence on prognostic prediction model.

We constructed two Cox regression models, Coxm1 and Coxm2, using clinical features and nutritional indicators from the training cohort of patients with digestive system cancers. The Coxm1 model contained seven features, while Coxm2 incorporated an extra L3 CLSFV measured by CT. Performance was evaluated using multiple metrics in the validation cohort, including time-dependent receiver operating characteristic (timeROC), time-dependent concordance index (timeC-index), calibration curve, and Kaplan-Meier curve. The predictive accuracy of the model was further assessed using net reclassification improvement (NRI) and integrated discrimination improvement (IDI).

Both models had high area under the curve (AUC) (range, 0.78-0.89) and C-index values (range, 0.74-0.80). The timeROC curve showed that the inclusion of L3 CLSFV in Coxm2 did not improve the model's AUC, with similar values observed at 1-, 3-, and 5-year. Coxm2 did not improve time-dependent C-index in comparison with Coxm1. Calibration curves showed good agreement between predicted and actual survival probability in both models, with slight improvements seen in Coxm2 versus Coxm1 (Brier score, 0.166 vs. 0.168). NRI indicated that the inclusion of L3 CLSFV in Coxm2 improved the model's performance. The category NRI of Coxm2 versus Coxm1 was 0.0768 [95% confidence interval (CI): -0.0768 to 0.128], while the continuous NRI was 0.0639 (95% CI: -0.0363 to 0.293). The IDI of Coxm2 versus Coxm1 was 0.006 (95% CI: -0.003 to 0.028). In the Kaplan-Meier curves, both Coxm1 and Coxm2 were accurately differentiated between high- and low-risk groups.

The addition of the L3 CLSFV to the Cox model improved the predictive accuracy and reclassification ability. These findings suggest that incorporating extra nutritional indicators can enhance the performance of prognostic models in digestive system cancer.