sEBM: Scaling Event Based Models to Predict Disease Progression via Implicit Biomarker Selection and Clustering

Area of Science:

• Computational Biology and Bioinformatics
• Biostatistics and Data Science
• Neuroscience and Neurodegenerative Diseases

Background:

• The Event Based Model (EBM) is a probabilistic generative model for disease progression via biomarker events.
• Traditional EBM struggles with high-dimensional data due to the combinatorial explosion of possible event sequences.
• Markov Chain Monte-Carlo (MCMC) is used for posterior distribution sampling, but becomes computationally intractable with many biomarkers.

Purpose of the Study:

• To develop a scalable Event Based Model (sEBM) for analyzing high-dimensional biomarker data in disease progression.
• To implicitly select relevant biomarkers and cluster them to reduce the complexity of event sequence inference.
• To validate sEBM's performance on synthetic and real-world clinical data, including Alzheimer's Disease progression.

Main Methods:

• The scaled Event Based Model (sEBM) implicitly selects a subset of biomarkers for event sequence inference.
• sEBM clusters biomarkers with similar event sequence positions, ordering clusters instead of individual biomarkers.
• The method was validated using synthetic data and applied to Alzheimer's Disease Neuroimaging Initiative (ADNI) data.

Main Results:

• sEBM significantly reduces the space of possible event sequences, enabling analysis with up to 300 biomarkers.
• On synthetic data, sEBM outperforms previous EBM variants in higher dimensional settings.
• Application to ADNI data successfully stratified subjects (cognitively normal, MCI, AD) into 6 disease stages.
• Increased sEBM stage strongly predicts conversion risk to Alzheimer's Disease in MCI subjects.

Conclusions:

• The scaled Event Based Model (sEBM) effectively addresses the scalability limitations of traditional EBM for high-dimensional biomarker data.
• sEBM provides a robust framework for modeling disease progression and stratifying patients without relying on a priori diagnostic labels.
• The method demonstrates significant potential for clinical applications, particularly in predicting disease progression and conversion risk in neurodegenerative diseases.