Machine learning-based ultrasomics for predicting response to tyrosine kinase inhibitor in combination with anti-PD-1 antibody immunotherapy in hepatocellular carcinoma: a two-center study.

Objective: The objective of this study is to build and verify the performance of machine learning-based ultrasomics in predicting the objective response to combination therapy involving a tyrosine kinase inhibitor (TKI) and anti-PD-1 antibody for individuals with unresectable hepatocellular carcinoma (HCC). Radiomic features can reflect the internal heterogeneity of the tumor and changes in its microenvironment. These features are closely related to pathological changes observed in histology, such as cellular necrosis and fibrosis, providing crucial non-invasive biomarkers to predict patient treatment response and prognosis.

Methods: Clinical, pathological, and pre-treatment ultrasound image data of 134 patients with recurrent unresectable or advanced HCC who treated with a combination of TKI and anti-PD-1 antibody therapy at Henan Provincial People's Hospital and the First Affiliated Hospital of Zhengzhou University between December 2019 and November 2023 were collected and retrospectively analyzed. Using stratified random sampling, patients from the two hospitals were assigned to training cohort ( = 93) and validation cohort ( = 41) at a 7:3 ratio. After preprocessing the ultrasound images, regions of interest (ROIs) were delineated. Ultrasomic features were extracted from the images for dimensionality reduction and feature selection. By utilizing the extreme gradient boosting (XGBoost) algorithm, three models were developed: a clinical model, an ultrasomic model, and a combined model. By analyzing the area under the receiver operating characteristic (ROC) curve (AUC), specificity, sensitivity, and accuracy, the predicted performance of the models was evaluated. In addition, we identified the optimal cutoff for the radiomic score using the Youden index and applied it to stratify patients. The Kaplan-Meier (KM) survival curves were used to examine differences in progression-free survival (PFS) between the two groups.

Results: Twenty ultrasomic features were selected for the construction of the ultrasomic model. The AUC of the ultrasomic model for the training cohort and validation cohort were 0.999 (95%CI: 0.997-1.000) and 0.828 (95%CI: 0.690-0.966), which compared significant favorably to those of the clinical model [AUC = 0.876 (95%CI: 0.815-0.936) for the training cohort, 0.766 (95%CI: 0.597-0.935) for the validation cohort]. Compared to the ultrasomic model, the combined model demonstrated comparable performance within the training cohort (AUC = 0.977, 95%CI: 0.957-0.998) but higher performance in the validation cohort (AUC = 0.881, 95%CI: 0.758-1.000). However, there was no statistically significant difference ( > 0.05). Furthermore, ultrasomic features were associated with PFS, which was significantly different between patients with radiomic scores (Rad-score) greater than 0.057 and those with Rad-score less than 0.057 in both the training (HR = 0.488, 95% CI: 0.299-0.796, = 0.003) and validation cohorts (HR = 0.451, 95% CI: 0.229-0.887, = 0.02).

Conclusion: The ultrasomic features demonstrates excellent performance in accurately predicting the objective response to TKI in combination with anti-PD-1 antibody immunotherapy among patients with unresectable or advanced HCC.