Multimodal deep learning approaches for improving polygenic risk scores with imaging data

Enhancing the predictive accuracy of polygenic risk scores (PRS) can be achieved by incorporating imaging data, which captures disease-related phenotypic variations. This study aimed to evaluate the predictive performance of PRS, image-derived deep learning scores (IDS), imaging-derived phenotypes (IDPs), and clinical covariates. Additionally, IDP-proxy genetic scores (IGSs) were developed as substitutes for IDPs in cases where imaging data are unavailable. The study assessed the predictive performance of these variables for glaucoma using the XGBoost model. IDS was generated using a deep learning model trained on optical coherence tomography (OCT) images obtained from the UK Biobank. The dataset included 55,469 samples with OCT imaging data and 402,847 samples without OCT images. To accommodate cases without imaging data, IGSs were introduced as replacements for IDPs. The AUC was used to compare the performance of individual predictors and multimodal models. Among single predictors, IDS demonstrated the highest predictive performance, achieving an AUC of 0.7742, indicating its effectiveness in capturing glaucoma-related structural changes. IDPs and PRS showed lower AUCs of 0.7106 and 0.6150, respectively. The multimodal model integrating PRS, IDS, IDPs, and clinical covariates achieved the highest predictive accuracy with an AUC of 0.7921. When imaging data were unavailable, IGSs alone had a relatively low AUC (0.5467), but performance improved to 0.7397 when combined with PRS and clinical covariates, demonstrating their viability as substitutes for IDPs. This study highlights the effectiveness of deep learning-based image features in glaucoma risk prediction and demonstrates that integrating genetic, imaging, and clinical data enhances predictive accuracy. Additionally, IGSs provide a viable alternative to IDPs when imaging data are unavailable, ensuring reliable glaucoma risk assessment in resource-limited settings.