In clinical gait analysis, the gait of a patient is recorded with optical motion capture and compared with a healthy reference group. High-dimensional gait datasets are difficult to interpret; machine learning can provide guidance regarding the most relevant gait phases and joint angles for visual analysis and quantify the difference between healthy and pathological gait. We propose an explicit state duration hidden Markov model (HMM) modeling the timeseries data of a subject or a group and the use of a reference-based measure that compares the most likely observations in each state. Based on this stochastic framework, the similarity between healthy and pathological gait can be quantified for each state, each joint angle, and each subject. This concept also includes an overall gait index useful for group comparison or the assessment of an individual's gait. For visualization, joint angle timeseries can be generated from the explicit state duration HMM. The accuracy of the explicit state duration HMM and the performance of the reference-based measures are evaluated on a dataset including strides of healthy subjects and patients suffering from arthritis.