Abstract
Accurate dementia risk prediction is challenging, and may be facilitated by better use of imaging and genetic data, including their complex interactions. We explored using deep survival neural networks to integrate these multi-modal, high-dimensional data. We included 3521 Rotterdam Study participants, 6340 magnetic resonance imaging (MRI) scans, with follow-up clinical diagnosis for dementia, and used 504 samples from Alzheimer's Disease Neuroimaging Initiative (ADNI) as an external validation. Genetic data included APOE-ε4 status and 76 additional SNPs. We developed models combining Convolutional Neural Networks (CNN) and Cox Proportional Hazards (CPH) models and provided post-hoc explanations. Our models outperformed CPH models including age, sex, and genetic inputs in both Rotterdam Study and ADNI by C-index of 0.88/0.63 V.S. 0.85/0.58, p-value of 0.02/0.002. Although their performance did not surpass CPH models also included MRI markers (0.89/0.66), additional predictability was obtained in age-stratified prediction in ADNI. Incorporating CNN image features in CPH models further increased performance to highest C-index of 0.90/0.69. Age and image had the highest importance in prediction, with age, image and genetic features showing the strongest interactions. Our approach indicates that imaging and genetic data can be feasibly integrated for dementia risk prediction, with informative extraction, reliable explanations and potential predictive gains.