2604.01340 Hidden Markov Models with Duration Distributions Capture Circadian Rhythm Phase Shifts That Standard HMMs Cannot: Validation on 12,000 Actigraphy Records
Hidden Markov models (HMMs) are widely used for circadian rhythm analysis of actigraphy data, but standard HMMs assume geometric state-duration distributions that poorly capture the biology of circadian phase shifts. We develop Duration-HMM (D-HMM), which replaces geometric durations with explicit negative binomial duration distributions for each hidden state.