Disease Progression Modeling and Stratification for detecting sub-trajectories in the natural history of pathologies: Application to Alzheimer’s disease trajectory modeling

Quantifying the progression of degenerative diseases remains crucial for early diagnosis, prevention, and treatment. However, accurately modeling disease biomarker evolution is hindered by substantial variability in disease trajectories among individuals, driven by demographic, genetic, and lifestyle factors. This variability gives rise to heterogeneous phenotypic manifestations, underscoring the need for stratification based on underlying disease subtypes. Recent advances have shown promise in unsupervised stratification of disease trajectories. Yet, current approaches face significant challenges related to robustness, biomarker specificity, interpretability, and temporal resolution of clustering results. To address these challenges, we introduce Disease Progression Modeling and Stratification (DPMoSt), a new probabilistic model designed to optimize clusters of continuous trajectories along a long-term disease time axis. This approach allows for the determination of subtype-specific biomarkers, improving the accuracy of patient stratification and generalization on external cohorts. We demonstrate DPMoSt on both synthetic and real-world data for the modeling of Alzheimer's disease (AD) evolution. In the synthetic experiments, DPMoSt shows high accuracy in reconstructing trajectory subtypes and identifying the biomarkers' specificity for the clustering problem. Our experiments in the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate the ability of DPMoSt to identify AD subtypes associated with accelerated cognitive decline and higher prevalence of the APOE4 variants. This result was validated on the external memory clinic cohort of the Geneva University Hospitals, confirming the association between cognitive decline and APOE4 in the pathological subtype. These results highlight the robustness of DPMoSt as well as its potential for broader applicability, offering a powerful tool for studying disease progression and subtype differentiation across diverse populations.