Graph Neural Networks for HCC Risk Adjustment and Interoperability

Hepatocellular carcinoma (HCC) is the leading cause of cancer-related death among patients with

cirrhosis. Accurate estimation of prognostic risk is crucial for patient management and clinical trial design. Existing

models for risk prediction after diagnosis of HCC are affected by high degrees of uncertainty, particularly in children

and older adults. Additionally, while substantial heterogeneity exists in clinical outcomes, risk-prediction models do not

officially incorporate it. One potential way to address these issues is through outcome predictive modeling techniques

such as survival analysis with random-effects. However, existing random-effects models for HCC have only been built

with a small number of variables, limiting translation into clinical practice. Two data sources, the National Inpatient

Sample and the United Network for Organ Sharing dataset, were therefore merged.

Data quality and heterogeneity remain major challenges and limitations of these datasets. As in many domains, hospital

diagnosis codes are the only data readily available for a substantial share of patients. Furthermore, different

organizations use different combinations of variables for clinical assessment and prediction. The use of data from

multiple sources adds complexity because of differences in coding and outcomes between organizations. Model

performance will not be meaningfully improved by simply using more variables. Standard summary statistics on

prediction accuracy are sensitive to differences in data quality and coding practices. The utility of existing data

therefore extends beyond just risk prediction to risk adjustment, that is, ensuring fair comparisons across different demographics. An unsupervised graph representation of the data that jointly identified high-quality codes and dislikelike patterns was applied.