Comparing machine learning and logistic regression methods for predicting hypertension using a combination of gene...

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Genetic association studies face challenges due to high dimensionality (many variables relative to observations).
• Established best practices for applying machine learning (ML) in genetic analysis are limited.
• Supervised ML offers potential for disease risk prediction using genetic data.

Purpose of the Study:

• To extend a supervised ML approach for disease risk prediction by incorporating gene expression data and rare variants.
• To evaluate the performance of radial and linear support vector machines (SVMs) against logistic regression using simulated genetic data.
• To assess the impact of including additional genes, including those with causal rare variants, on model predictive ability.

Main Methods:

• Application of extended supervised ML approach (radial and linear SVMs) to simulated data from Genetic Analysis Workshop 19.
• Comparison of ML methods' predictive performance against traditional logistic regression.
• Evaluation of model robustness with increasing numbers of genes, including causal rare variants.

Main Results:

• Overall method performance was comparable, with linear SVMs showing slight predictive gains over radial SVMs and logistic regression.
• Increasing the number of genes in models, even with causal rare variants, significantly decreased predictive ability for radial SVMs and logistic regression.
• Linear SVMs exhibited more robust performance when additional genes were included in the models.

Conclusions:

• Linear SVMs offer a more stable approach for disease risk prediction in genetic association studies compared to radial SVMs and logistic regression, particularly when dealing with numerous genes.
• Further research is required to validate ML approaches on larger datasets and quantify the benefits of integrating gene expression data for improved predictive modeling.