Beyond the Last Scan: The Role of Longitudinal Neuroimaging in Predicting Conversion to Alzheimer's Disease at Different Stages

The application of deep learning models to neuroimaging data has proven to be a powerful approach for understanding and predicting the progression of Alzheimer's disease (AD). A critical factor in improving prediction accuracy lies in in-corporating longitudinal information into these models. Previous approaches integrated longitudinal neuroimaging data from time steps in the future, and evaluated their re-sults on a single heterogeneous group of MCI subjects, con-taining MCI converters (MCI-C) at different stages of disease progression. In this study, we investigate the impact of incorporating past longitudinal neuroimaging information into deep learning models to predict the conversion from mild cognitive im-pairment (MCI) to Alzheimer's disease, and we evaluate our results for different MCI groups with converters at different stages of the disease. We demonstrate that incorporating past temporal patterns significantly enhances predictive performance, particularly in identifying early and intermediate stages of the disease. Early detection improved the area under the curve (AUC) by 8,1% while intermediate detection improved by 6.4%, compared with models that did not incorporate longitudinal data.