A complete data set of every movement of all the inpatients from room to room covering two years was provided us by the Medical Information Department of the University of Tsukuba Hospital in Japan. By focusing on the obstetric patients, who are assumed to be hospitalized rather at random times, we have analyzed the patient flow using our original visualization software. Upon admission, each obstetric patient is assigned to a bed in one of the two wards, one for high-risk delivery and the other for normal delivery, and then she may be transferred between the two wards before discharge. We confirm Little's law of queuing theory for the patient flow in each ward. Then we propose a network model of M/G/f f and M/M/m queues to represent the flow of these patients, which is used to predict the probability distribution for the number of patients staying in each ward at the nightly census time from the observed data of patient admission rate and the histogram of the length-of-stay (LOS) in that ward. Although our model is a very rough and simplistic approximation of the real patient flow, the predicted probability distribution is shown to be in good agreement with the observed one. Our method can be used for planning the capacity of obstetric units when the patient demand is predicted.