DATA-DRIVEN MODELS OF DISEASE PROGRESSION AND APPLICATIONS TO ALZHEIMER’S DISEASE: EVENT-BASED MODEL AND DIFFERENTIAL EQUATION MODELS OF BIOMARKER CHANGES IN ADNI

Abstract

Understanding and identifying early pathological changes of neurological diseases is key to realizing disease-modifying treatments, but identifying presymptomatic individuals is challenging. Data-driven disease progression modelling is a family of emerging techniques for achieving this by learning directly from measured data. We considered two such models. The event-based model1,2 (EBM) robustly finds the order in which a set of biomarkers become abnormal, using cross-sectional data, without defined biomarker cut-points, and requiring no prior clinical staging variable. This uniquely fine-grained picture of disease progression is capable of probabilistic, high-resolution patient staging and stratification, useful for clinical trials selection and end-points. However, the EBM doesn't quantify time. Differential equation models3,4 (DEMs) use short-term longitudinal change to find biomarker trajectories that can be used to stage patients temporally, and to predict symptom onset. Both approaches assume monotonic disease progression, and have succeeded in building data-driven models of neurodegenerative disease progression. We present and compare models built on the latest contents of the ADNI data set. Biomarker data is from 1737 ADNI-1/GO/2 participants (417 cognitively normal (CN), 106 significant memory concern (SMC), 310 early mild cognitive impairment (EMCI), 562 late MCI (LMCI), 342 diagnosed Alzheimer's disease (AD)): a multi-modal set of measurements derived from magnetic resonance imaging, positron emission tomography, cerebrospinal fluid, and cognitive test scores. From baseline data we estimated a pathological cascade in ApoE4-positive participants using an EBM. From short-term longitudinal data we estimated probabilistic biomarker trajectories for ApoE4-positive clinical progressors using nonparametric DEMs. The EBM-estimated AD pathological cascade (Figure 1) starts with accumulation of amyloid and tau, followed by cognitive abnormalities and temporal lobe volume reduction, then brain hypometabolism, and broader neurodegeneration. A similar pattern was found by DEMs (Figure 2), with biomarker trajectories accelerating from normal to abnormal levels. Patient staging with both models (Figures 3, 4) provided much finer resolution of AD progression than traditional diagnostic categorization, while remaining consistent with clinical diagnoses. We found a data-driven sequence, and timings, of AD biomarker progression in the ADNI dataset. Our data-driven model-based approaches increase understanding of neurological disease progression with potential utility for patient staging and prognosis. Event-based model of AD progression: probabilistic sequence of biomarker abnormality. Biomarkers (imaging/molecular/cognitive) along the vertical axis are ordered by the maximum likelihood disease progression sequence (from top to bottom). The horizontal axis shows estimated variance in the posterior sequence, with positional likelihood per event given by grayscale intensity. Differential equation model of AD progression: preclinical cumulative probability of biomarker abnormality, increasing from left (dark blue) to right (light yellow) as a function of time to AD diagnosis. Biomarker trajectories estimated from differential data are aligned at the average point of AD diagnosis. Biomarkers along the vertical axis (top to bottom) are ordered by the time where the cumulative abnormality probability equals 50%. Box plots of the interquartile range show the speed of abnormality progression for each biomarker. Event-based model fine-grained probabilistic staging of individuals across clinical groups in ADNI-1/GO/2. Differential equation model probabilistic temporal staging of individuals across clinical groups in ADNI-1/GO/2.