This study investigates the use of machine learning models to analyze liver cancer (HCC) recurrence after surgery in patients at NTUH between January 2001 and December 2012 (N = 2314). We explore two main outcomes:
- Early Recurrence (Binary Classification): Recurrence within 2 years post-surgery (only patients followed-up for more than 2 years were included).
- Recurrence-Free Time (Time-to-Event Outcome): Recurrence status and time elapsed before recurrence.
Time-to-Event Regression Models:
We employed two models for survival analysis:
- Cox Proportional Hazards Model (CoxPH): A widely used model assuming a constant effect of explanatory variables on the hazard rate.
- Random Survival Forests Model (RSF): An adaptation of random forests for survival analysis that handles right-censored data.
Key Findings:
- Early Recurrence Prediction: Random Forest exhibited the best performance for predicting recurrence within 2 years (AUC: 0.689 ± 0.013).
- Recurrence-Free Time Estimation: RSF demonstrated superior performance over the Cox model for estimating recurrence-free time. Feature selection further improved model performance. Early model performance was lower due to potentially incomplete tumor removal in some patients.
- Factors Associated with Recurrence: The Cox model identified several factors associated with increased risk of recurrence (HR > 1, p-value < 0.05):
- Tumor size
- ALBI grade (liver function)
- Satellite nodules
- Cirrhosis
- Surgical margin
- Tumor stage
- Age
slides: https://docs.google.com/presentation/d/1FL32_nfl9ia0vp-YJKspW2qhHeIGWtcDIz1tsaKbTeU/edit?usp=sharing