Efficient Learning of Continuous-Time Hidden Markov Models for Disease Progression.

Area of Science:

• Biostatistics
• Computational Biology
• Machine Learning

Background:

• Continuous-Time Hidden Markov Models (CT-HMMs) are valuable for modeling irregular, noisy disease progression data.
• Existing parameter learning algorithms for CT-HMMs are inefficient, limiting their application to small models or requiring restrictive assumptions.

Purpose of the Study:

• To present the first comprehensive characterization of efficient EM-based learning methods for CT-HMMs.
• To overcome the computational limitations hindering the use of complex CT-HMMs in disease progression modeling.

Main Methods:

• Reformulated state probability estimation using an equivalent discrete time-inhomogeneous hidden Markov model.
• Adapted three continuous-time Markov chain approaches for end-state conditioned statistics computation in CT-HMMs.
• Developed and applied efficient EM-based learning algorithms for CT-HMM parameter estimation.

Main Results:

• Successfully addressed the dual challenges of posterior state probability estimation and end-state conditioned statistics computation.
• Enabled the use of CT-HMMs with over 100 states, significantly expanding model complexity.
• Demonstrated effective disease progression modeling and prediction using glaucoma and Alzheimer's disease datasets.

Conclusions:

• The developed EM-based learning methods provide an efficient solution for CT-HMM parameter learning.
• This advancement facilitates the application of complex CT-HMMs for robust disease progression analysis and prediction.
• The approach shows promise for improving understanding and management of neurodegenerative and ocular diseases.