Parsimonious EBM: Generalising the event-based model of disease progression for simultaneous events

Abstract

The event-based model of disease progression (EBM) infers a temporal ordering of biomarker abnormalities, defining different disease stages, from cross-sectional data. A key modelling choice of the EBM is that biomarker abnormalities, termed events, are serially ordered. However, this choice enforces a strict equality between the number of input biomarkers and the number of modelled disease stages, limiting the EBM's ability to infer simple staging systems and identify latent disease processes driving multiple biomarker changes. To overcome this, we introduce the parsimonious event-based model of disease progression (P-EBM). The P-EBM generalises the EBM to allow multiple new biomarker abnormalities, termed "simultaneous events", at each model stage. We evaluate the P-EBM performance in simulated data and demonstrate its ability to reconstruct event orderings with arbitrary arrangements under realistic experimental conditions. In sporadic AD data from the Alzheimer's Disease Neuroimaging Initiative, the P-EBM estimated a sequence with 7 model stages from a dataset of 12 biomarkers that more closely fitted the data than the EBM. The inferred sets of simultaneous events, such as decreased cerebrospinal fluid total tau and p-tau₁₈₁, correspond closely to known underlying disease processes. P-EBM patient stages were strongly associated with clinical diagnosis at baseline and future conversion and could be accurately estimated from a smaller number of biomarkers than the EBM. The P-EBM enables the data-driven discovery of simple disease staging systems which could highlight new latent disease processes and suggest practical strategies for patient staging.