Investigating the Use of Machine Learning Methods to Build Risk Prediction Models for Complex Disease

摘要

Large-scale population biobanks offer exciting opportunities to develop risk prediction models for complex diseases because of the availability of genetic data and extensive lifestyle and clinical information. Unlike traditional polygenic risk scores, machine learning methods can be utilized to build risk prediction models that include both genetic and non-genetic features, and interactions between them. We have performed a simulation study to assess the utility of several machine learning methods (gradient boosting machines, deep learning neural networks, and random forests) to generate prediction models for type 2 diabetes (T2D), applied using the H2O package, using data from the UK Biobank. Twenty thousand participants were randomly selected according to their T2D status (10,000 cases and 10,000 controls). Five relevant clinical factors (age, sex, body mass index, diastolic blood pressure and systolic blood pressure) were selected for entry into the model alongside a set of 1-100 SNPs, simulated with varying minor allele frequency and relative risk of disease. Irrelevant clinical factors were also selected to assess whether the methods identify them as unimportant for disease prediction. All methods successfully identified the most strongly associated genetic and non-genetic factors as the most important features for prediction, and assigned the least importance to the irrelevant factors. Results also indicated that the inclusion of strongly associated genetic variants increases the predictive accuracy of the model compared to using clinical factors alone, while the inclusion of more modestly associated variants does not appear to improve predictive power.