Development and interpretation of a pathomics-driven ensemble predictive model for prognosis of intrahepatic cholangiocarcinoma.

BACKGROUND AND AIMS: Pathomics exhibited great potential in improving clinical management but lacked of pathomics model for the prognosis of intrahepatic cholangiocarcinoma (iCCA) because of limited interpretability. This study aimed to develop and interpret a pathomics ensemble model for the prognosis of iCCA.

METHODS

Based on 252 iCCA patients in our center, we constructed a deep learning framework to extract pathomics features from the postoperative tumor slide, which was externally validated in The Cancer Genome Atlas database. An ensemble predictive model was then developed based on the pathomics score and clinical parameters. The areas under the receiver operating characteristic curve (AUCs) were calculated to evaluate the performances. K-means unsupervised clustering classifies iCCA into subtypes based on the pathomics features. Cellprofiler was employed to extract the quantitative histomorphological features, and the differences between subtypes in clinical outcomes, histomorphology features, and genomics were further investigated.

RESULTS

The AUCs of the ensemble pathomics model for the 1-year, 2-year, and 3-year OS reached 0.902, 0.853, and 0.807, respectively. The Gradient-weighted class activation mapping analysis was performed to visualize the model attention area. Two clusters were identified using the K-means clustering, which exhibited significantly different prognosis and histomorphological features. Several genes were significantly differentially expressed, several pathways were highly enriched, and two prognosis-related genes (TPM2 and PLOD2) were identified.

CONCLUSIONS

This study initially developed an ensemble pathomics model for the prognosis of iCCA, which could help make individualized treatment postoperatively. Moreover, we employed various methods to improve the interpretation of the model.