Machine learning: an approach to preoperatively predict PD-1/PD-L1 expression and outcome in intrahepatic cholangiocarcinoma using MRI biomarkers

Objective: To investigate the preoperative predictive value of non-invasive imaging biomarkers for programmed cell death protein 1/programmed cell death protein ligand 1 (PD-1/PD-L1) expression and outcome in intrahepatic cholangiocarcinoma (ICC) using machine learning.

Methods: PD-1/PD-L1 expression in 98 ICC patients was assessed by immunohistochemistry, and their prognostic effects were analysed using Cox regression and Kaplan-Meier analysis. Radiomic features were extracted from MRI in the arterial and portal vein phases, and three sets of Radiomics score (Radscore) with good performance were derived respectively as biomarkers for predicting PD-1, PD-L1 expression and overall survival (OS). PD-1 and PD-L1 expression models were developed using the Radscore (arterial phase), clinico-radiological factors and clinical factors, individually and in combination. The imaging-based OS predictive model was constructed by combining independent predictors among clinico-radiological, clinical factors and OS Radscore. Pathology-based OS model using pathological and clinical factors was also constructed and compared with imaging-based OS model.

Results: The highest area under the curves of the models predicting PD-1 and PD-L1 expression was 0.897 and 0.890, respectively. PD-1⁺ and PD-L1⁺ cases had worse outcomes than negative cases. The 5-year survival rates of PD-1⁺ and PD-1^- cases were 12.5% and 48.3%, respectively (p<0.05), whereas the 5-year survival was 21.9% and 39.4% for PD-L1⁺ and PD-L1^- cases, respectively (p<0.05). The imaging-based OS model involved predictors of clinico-radiological 'imaging classification', radiomics 'Radscore' from arterial phase and carcinoembryonic antigen (CEA) level (C-index:0.721). It performed better than pathology-based model (C-index: 0.698) constructed by PD-1/PD-L1 expression status and CEA level. The imaging-based OS model is potential for practice when the pathology assay is unavailable and could divide ICC patients into high-risk and low-risk groups, with 1-year, 3-year and 5-year survival rates of 57.1%, 14.3% and 12.4%, and 87.8%, 63.3% and 55.3%, respectively (p<0.001).

Conclusions: MRI radiomics could derive promising and non-invasive biomarker in evaluating PD-1/PD-L1 expression and prognosis of ICC patients.