ABSTRACTA patient’s risk for cancer is usually estimated through simple linear models that sum effect sizes of proven risk factors. In theory, more advanced machine learning models can be used for the same task. Using data from the UK Biobank, a large prospective health study, we have developed linear and machine learning models for the prediction of 12 different cancers diagnoses within a 10 year time span. We find that the top machine learning algorithm, XGBoost (XGB), trained on 707 features generated an average area under the receiver operator curve of 0.736 (with a range of 0.65-0.85). Linear models trained with only 10 features were found to be statistically indifferent from the machine learning performance. The linear models were significantly more accurate than the prominent QCancer models (p = 0.0019), which are trained on 45 million patient records and available to over 4,000 United Kingdom general practices. The increase in accuracy may be caused by the consideration of often omitted feature types, including survey answers, census records, and genetic information. This approach led to the discovery of significant novel risk features, including self-reported happiness with own health (relevant to 12 cancers), measured testosterone (relevant to 8 cancers), and ICD codes for rehabilitation procedures (relevant to 3 cancers). These ten feature models can be easily implemented within the clinic, allowing for personalized screening schedules that may increase the cancer survival within a population.
Machine Learning–enabled Scalable Performance Prediction of Scientific Codes