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

Area of Science:

• * Medical Informatics
• * Machine Learning in Oncology
• * Radiation Oncology

Background:

• * Federated learning (FL) enables multi-institutional predictive modeling without sharing raw patient data, preserving privacy.
• * Linear FL (LFL) offers an interpretable approach to enhance sample size and generalizability in outcome prediction.
• * Hepatocellular carcinoma (HCC) patients undergoing external beam radiotherapy (EBRT) are a key focus for predictive modeling.

Purpose of the Study:

• * Evaluate LFL for predicting hepatic toxicity and 1-year survival (SRVy1) in HCC patients.
• * Assess LFL's ability to enhance generalizability and interpretability in multi-institutional settings.
• * Demonstrate LFL as a proof-of-concept for privacy-preserving collaborative research.

Main Methods:

• * Patient data from Massachusetts General Hospital (MGH) and Brigham and Women's Hospital (BWH) for training; MD Anderson Cancer Center for validation.
• * Logistic regression models predicting hepatic toxicity and SRVy1 using clinical features (albumin, bilirubin, Child-Pugh score, liver size, mean liver dose).
• * Local model training at each institution with LFL, without raw data sharing; performance evaluated using AUC.

Main Results:

• * LFL improved survival prediction AUC from 0.55-0.63 (single-institution) to 0.67.
• * LFL maintained AUC at 0.7 for toxicity prediction, comparable to single-institution models (AUC 0.68-0.69).
• * LFL demonstrated moderate coefficients, mitigating bias and improving external validation performance.

Conclusions:

• * LFL maintained or improved predictive performance for survival and hepatic toxicity in HCC patients.
• * LFL preserves model interpretability and patient privacy in multi-institutional collaborations.
• * Findings support LFL's utility in advancing collaborative research in radiation oncology.