Novel machine learning algorithm in risk prediction model for pan-cancer risk: application in a large prospective...

Area of Science:

• Oncology
• Bioinformatics
• Public Health

Background:

• Pan-cancer incidence prediction is crucial for early detection and prevention strategies.
• Leveraging routine health check-up data offers a scalable approach for risk assessment in large populations.

Purpose of the Study:

• To develop and validate machine-learning models for predicting pan-cancer incidence risk.
• To utilize demographic, questionnaire, and routine health check-up data in a large Asian cohort.

Main Methods:

• A prospective cohort study of 433,549 participants (male and female cohorts) over an 8-year median follow-up.
• Development and comparison of machine learning models including XGBoost, Lasso-Cox, and Random Survival Forests.
• Evaluation of model performance using Area Under the Curve (AUC) and identification of key predictive variables.

Main Results:

• XGBoost demonstrated superior performance in predicting pan-cancer incidence compared to other models.
• Optimized models with fewer variables (31 for males, 11 for females) maintained high predictive accuracy (AUCs ranging from 0.746 to 0.876).
• High-risk individuals identified by the models exhibited at least a ninefold higher risk of pan-cancer incidence.

Conclusions:

• The study successfully developed and internally validated machine-learning models for pan-cancer risk prediction using routine health data.
• These models demonstrate good discriminatory ability with a parsimonious set of predictors.
• External validation is recommended prior to clinical implementation of the risk prediction model.