Using deep learning to improve genetic studies of osteoporosis
Eriksson, T.; Nakamori, C.
Show abstract
To evaluate how recent advances in deep learning can improve the construction of quantitative phenotypes for genome-wide association studies (GWAS), we focused on the context of osteoporosis and bone mineral density (BMD) measurements. We applied image classifiers and transformer models to three distinct tasks. First, we developed quantitative estimates of osteoporosis severity using bone X-ray images. Second, we compared standard approaches for handling confounding variables with a multi-factor strategy based on transformer models trained on UK Biobank data. Third, we investigated whether image-based models could predict how single nucleotide polymorphisms (SNPs) associated with BMD influence bone structure. While our results were promising, application of deep learning methods did not yield substantial improvements over established approaches. Nonetheless, our findings highlight the potential of integrating imaging and machine learning techniques to refine phenotype definitions in genetic studies.
Matching journals
The top 1 journal accounts for 50% of the predicted probability mass.