Machine learning-based multimodal prediction of prognosis in patients with resected intrahepatic cholangiocarcinoma.

Abstract

4121 Background: Intrahepatic cholangiocarcinoma (iCCA) is an aggressive malignancy and the second most common primary liver cancer. About a third of patients may benefit from surgery but recurrence is common and the overall survival is low. Accurate prognosis modeling that can predict response to treatment and outcome remains an unmet clinical need. Deep learning methods provide a new opportunity to better predict prognosis by extracting distinguishing characteristics interrogating multiple data sources, which appears to be superior to unimodal modeling. This study aimed to construct a predictive model of patient outcome and identify predictive biomarkers based on a combination of clinical, genomic, histological and radiological data. Methods: We analyzed 83 patients with iCCA that underwent surgery and designed a multi-modal model to predict overall survival (OS) and progression-free survival (PFS). Among these patients, each had one to three hematoxylin/eosin histology slides. As a pre-processing step, each histology slide was decomposed into patches, each patch was fed into a ResNet feature extractor to extract a feature representation of the patch, and the features across patches were aggregated with maximum pooling to yield a 2048-dimensional vector representation for each slide. Among these 83 patients, 76 patients had their tumors profiled for somatic genomic alterations using MSK-IMPACT, a deep targeted-sequencing assay. A binary matrix across altered genes and samples was created as an input to the prediction model. Cox proportional hazard models specific to each modality (clinical data, histological slides, and altered genes) were then designed and their predictions were averaged to produce the final log-risk score. The models were validated using a 5-fold, 5-repeat cross-validation with patient-level splits. Results: A model using only clinical and routine histological data achieved a concordance index of 0.74 (95% confidence interval, 0.64-0.82) for OS and 0.73 (0.64-0.78) for PFS. Adding information about genetic alterations improved performance for OS (0.80, 0.70-0.88) and similar performance for PFS (0.72, 0.64-0.79). Both models outperform a staging-based patient stratification. Conclusions: This study demonstrates that machine learning models can improve survival prediction using multi-modal data after resection of intrahepatic cholangiocarcinoma. Such models have the potential to improve risk stratification and treatment recommendations. [Table: see text]