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

Area of Science:

• Ophthalmology
• Medical Imaging
• Genetics

Background:

• Polygenic risk scores (PRS) are limited in predicting disease risk.
• Imaging data capture phenotypic variations crucial for disease prediction.
• Glaucoma risk prediction can benefit from multimodal data integration.

Purpose of the Study:

• To evaluate the predictive performance of PRS, image-derived deep learning scores (IDS), imaging-derived phenotypes (IDPs), and clinical covariates for glaucoma.
• To develop and assess IDP-proxy genetic scores (IGSs) as substitutes for IDPs when imaging data are unavailable.
• To compare the predictive accuracy of individual predictors and multimodal models for glaucoma risk.

Main Methods:

• A deep learning model was trained on optical coherence tomography (OCT) images from the UK Biobank to generate IDS.
• XGBoost model was employed to assess predictive performance using AUC.
• IGSs were developed as genetic proxies for IDPs.

Main Results:

• IDS achieved the highest AUC (0.7742) among single predictors, outperforming IDPs (0.7106) and PRS (0.6150).
• The multimodal model integrating PRS, IDS, IDPs, and clinical covariates yielded the highest AUC (0.7921).
• IGSs combined with PRS and clinical covariates achieved an AUC of 0.7397, showing viability as IDP substitutes.

Conclusions:

• Deep learning-based image features are effective for glaucoma risk prediction.
• Integrating genetic, imaging, and clinical data significantly enhances predictive accuracy.
• IGSs offer a valuable alternative for glaucoma risk assessment in settings lacking imaging data.