Semi-supervised learning to improve generalizability of risk prediction models

Area of Science:

• Oncology
• Biostatistics
• Machine Learning in Healthcare

Background:

• Prediction model utility hinges on generalizability across diverse patient populations.
• Assessing generalizability is crucial for reliable clinical decision-making.

Purpose of the Study:

• To evaluate if semi-supervised learning improves colorectal cancer (CRC) risk prediction generalizability over supervised methods.
• To compare the discrimination, calibration, and clinical usefulness of semi-supervised versus supervised learning models for CRC survival risk.

Main Methods:

• Developed and validated a CRC survival risk prediction model using semi-supervised logistic regression.
• Utilized a large dataset (113,141 patients) from the Surveillance Epidemiology End Results registry for development.
• Tested generalizability on an independent cohort (1149 patients) from the Second Affiliated Hospital, Zhejiang University School of Medicine.

Main Results:

• All models demonstrated good discrimination.
• Supervised learning models showed poor calibration, while the semi-supervised model exhibited good calibration on the validation cohort, indicating superior generalizability.
• The semi-supervised logistic regression model led to better clinical outcomes in usefulness analysis.

Conclusions:

• Semi-supervised logistic regression offers improved generalizability for CRC risk prediction models.
• This approach enhances model calibration and clinical utility, providing a valuable reference for predictive model development in clinical practice.