Deep learning studies link MRI and genetics to dementia risk and Alzheimer's progression

Deep learning models using structural MRI and genetic data showed improved prediction of dementia risk and identified dynamic brain changes during the progression of Alzheimer’s disease. In one study, models integrating imaging and genetic data in 3,521 Rotterdam Study participants and 515 external validation samples achieved a highest C-index of 0.90/0.69, while a separate longitudinal MRI study reported that the importance of regions such as the amygdala, parahippocampal gyrus, and temporal lobe undergoes dynamic changes throughout AD progression.

Accurate dementia risk prediction is challenging, and may be facilitated by better use of imaging and genetic data, including their complex interactions. One study included 3,521 Rotterdam Study participants, 6,340 magnetic resonance imaging scans, with follow-up clinical diagnosis for dementia, and used 515 samples from the Alzheimer’s Disease Neuroimaging Initiative as an external validation. Genetic data included APOE-ε4 status and 76 additional SNPs. The researchers developed models combining Convolutional Neural Networks and Cox Proportional Hazards models and provided post-hoc explanations.

The models outperformed Cox Proportional Hazards models including age, sex, and genetic inputs in both the Rotterdam Study and external validation by C-index of 0.88/0.63 versus 0.85/0.58, with p-value of 0.02/0.002. Although their performance did not surpass Cox Proportional Hazards models that also included MRI markers (0.89/0.66), additional predictability was obtained in age-stratified prediction in external validation. Incorporating CNN image features in Cox Proportional Hazards models further increased performance to the 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.

A separate study said Alzheimer’s disease is an irreversible neurodegenerative disorder whose progression is closely associated with time, but that most diagnostic models are based on single time-point data, overlooking longitudinal disease characteristics. Structural magnetic resonance imaging has been widely utilized in the study of AD. To address the need for multi-time series analysis in longitudinal AD research and the integration of features from different brain tissues, the researchers proposed a Multi-Branch Fusion Channel Attention Network (MBFCA-Net) for disease diagnosis.

The network leverages the temporal correlations across longitudinal scans for effective AD detection. The study also conducted retrospective interpretability analysis to quantify the contributions of brain regions across disease stages. The results indicate that the importance of regions such as the amygdala, parahippocampal gyrus, and temporal lobe undergoes dynamic changes throughout the progression of AD. Furthermore, AD-related voxel clusters exhibit a developmental trend, shifting from the hippocampus to the temporal lobe and transitioning from a dispersed to a more aggregated distribution.

The studies said imaging and genetic data can be feasibly integrated for dementia risk prediction, with informative extraction, reliable explanations and potential predictive gains, and that longitudinal patterns of AD-related changes may offer contributions to early diagnosis and pathological understanding of the disease.