解釈可能なLASSO-Coxモデル：Hsp90α/アルブミン比による肝細胞がんの予後予測

The heat shock protein (HSP)/albumin (ALB) ratio, which measures the levels of plasma Hsp90α to ALB, is believed to be a sensitive prognostic biomarker for hepatocellular carcinoma (HCC); however, its prognostic value has not been clearly confirmed. The aim of our study was to assess the role of HSP/ALB as a prognostic factor in patients with HCC and further establish an interpretable predictive model using least absolute shrinkage and selection operator (LASSO)-Cox regression and SHapley Additive exPlanations (SHAP) value.

A total of 1,340 HCC patients were enrolled in this study. We measured the levels of HSP/ALB in HCC patients within a week before treatment. Then, we established the HSP/ALB ratio and assessed its correlation with prognosis. Finally, we use LASSO-Cox regression to incorporate HSP/ALB into the prediction model for visualization and risk stratification.

The best cut-off value for HSP/ALB was determined to be 3.77 using the X-tile software. It was found that the median overall survival (mOS) of patients with HSP/ALB ≥3.77 (high) [8.3 months, 95% confidence interval (CI): 7.6-9.3] was significantly lower than that of patients with HSP/ALB <3.77 (low) (38.7 months, 95% CI: 35.8-45). In the training cohort, the mOS of patients with HSP/ALB ≥3.77 was 7.3 months (95% CI: 5.8-8.9), and that of patients with HSP/ALB <3.77 was 36.1 months (95% CI: 28.4-39.8). In the validation cohort, the mOS of patients with HSP/ALB ≥3.77 was 7.1 months (95% CI: 4.9-9.0), and that of patients with HSP/ALB <3.77 was 36.2 months (95% CI: 31.6-43.7). The subgroup analysis included four groups: transarterial chemoembolization (TACE) alone, TACE combined with operation, and radiotherapy (RT) group. The results showed that the patients with HSP/ALB <3.77 had longer mOS. We integrated HSP/ALB into the LASSO-Cox prediction model. Our findings showed that using HSP/ALB alone, the area under the curve (AUC) values for predicting 1-, 2-, and 3-year overall survival (OS) were 0.757, 0.731, and 0.713, respectively. However, after incorporating HSP/ALB into the prediction model based on LASSO-Cox regression, the AUC values for predicting 1-, 2-, and 3-year OS were 0.850, 0.842, and 0.810, respectively. Additionally, a nomogram was created based on this improved prediction model. Meanwhile, it was observed that the OS of the low-risk group significantly outperformed that of the medium-risk and high-risk groups. Finally, we ranked the importance of variables according to the SHAP value.

The integration of HSP/ALB into the LASSO-Cox prediction model improves its predictive performance, indicating that HSP/ALB can serve as a reliable prognostic predictor.