Linear Federated Learning for Outcome Prediction With Application to Hepatocellular Carcinoma Radiotherapy

Federated learning (FL) enables multi-institutional predictive modeling without sharing raw patient data, preserving privacy while leveraging diverse data sets. This study evaluates the use of linear FL (LFL) as an interpretable approach to enhance sample size and generalizability in outcome prediction. As a proof of concept, we applied LFL to patients with hepatocellular carcinoma (HCC) undergoing external beam radiotherapy (EBRT), predicting hepatic toxicity and 1-year survival (SRVy1).

Patient data from Massachusetts General Hospital (MGH) and Brigham and Women's Hospital (BWH) were used to train models, whereas an independent validation data set from MD Anderson Cancer Center assessed generalizability. Logistic regression was developed to predict hepatic toxicity and SRVy1 using key clinical features, including baseline albumin, bilirubin, Child-Pugh score, liver size, and mean liver dose. The LFL approach allowed each institution to train models locally without sharing raw patient data. Model performance was evaluated using the AUC and compared between the LFL model and institution-specific models.

For survival prediction, single-institution models were limited, with AUC = 0.55-0.63, with LFL increasing it to 0.67. For toxicity prediction, external validation showed AUC = 0.68 for the MGH model and 0.69 for the BWH model, with LFL maintaining the AUC at 0.7. The model coefficients were moderate in the LFL compared with the single-institution models, indicating an ability to mitigate bias, which was also reflected by better performance on the validation data set.

LFL maintained or improved predictive performance over single-institution models for survival and hepatic toxicity in patients with HCC treated with EBRT while preserving model interpretability and patient privacy. These findings support LFL's role in multi-institutional collaborations.