FastEBM: Fast, Scalable, and Uncertainty-Aware Event-Based Disease Progression Modeling

Le résumé fourni par la source

Event-based models (EBMs) are used to infer ordering of biomarker alteration patterns with respect to disease progression. However, EBM approaches rely on computationally expensive permutation-based inference, assumptions of feature independence, and likelihood optimization that can limit scalability and stability in high-dimensional settings. Here, we introduce Fast Event-Based Model (FastEBM), a scalable, uncertainty aware, Markov- chain-based framework that reformulates disease progression inference as a subject-ordering problem on a data-driven diffusion manifold. The progression uncertainty, used to derive positional variance diagrams, is quantified using first-passage-time variability derived directly from the inferred Markov process. Using synthetic experiments varying feature dimensionality, cohort size, noise level, and feature-correlation structure, we compared FastEBM with established methods, including Gaussian mixture model EBM (GMM-EBM), kernel density estimation EBM (KDE-EBM), and discriminative EBM (DEBM). FastEBM achieved the best accuracy and runtime. In low-subject/high-dimensional stress tests, FastEBM retained event-order recovery. FastEBM remained robust in simulations containing correlated and redundant features after decorrelation and feature-group handling. We applied FastEBM to real-world data to characterize biomarker progression in Alzheimer's disease. First, we evaluated a low-dimensional multi- modal dataset from The Alzheimer's Disease Prediction Of Longitudinal Evolution (TAD- POLE) challenge. Second, to demonstrate high-dimensional disease progression mapping, we applied FastEBM to regional cortical tau-PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). In both cases, FastEBM recovered progression patterns broadly consistent with the literature, also revealing lateralized progression trends. These results show that diffusion-based Markov geometry provides a scalable and robust alternative to conventional event-based modeling. FastEBM is available at: https://github.com/sjusc07/FastEBM .

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.