Simple Linear Cancer Risk Prediction Models With Novel Features Outperform Complex Approaches.

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Accurate cancer risk prediction is crucial for precision prevention strategies.
• Current risk models, like the Gail model, are interpretable but may lack the accuracy of advanced machine learning (ML) models.

Purpose of the Study:

• To develop and assess predictive models for 13 cancer diagnoses within a 10-year timeframe using the UK Biobank prospective study.
• To compare the performance of machine learning (ML) models, linear models, and existing QCancer models for cancer risk prediction.

Main Methods:

• Developed ML and linear models utilizing all 931 features from the UK Biobank dataset.
• Assessed linear models trained on 10 selected features and externally validated QCancer models.
• Evaluated model performance using the area under the receiver operator curve (AUC).

Main Results:

• Linear models using all features achieved an average AUC of 0.722, slightly outperforming ML models (AUC 0.720).
• A 10-feature linear model demonstrated a comparable AUC (0.706) to models using all features and outperformed QCancer models (AUC 0.684).
• The high performance of the 10-feature linear model may be attributed to the inclusion of census and genetic data.

Conclusions:

• Unbiased selection of diverse features in linear models, rather than complex ML, can yield highly accurate cancer risk predictions.
• These findings suggest potential for personalized cancer screening schedules to enhance patient survival.
• The study highlights the importance of feature selection in developing effective predictive models for cancer.