According to the livestock statistics of the Ministry of Agriculture, Forestry, and Fisheries, the number of animals per household in Japan has increased since its peak in 1962; this indicates the recent shift in the trend toward large-scale farming in the livestock industry. Consequently, the occurrence of a livestock epidemic under such circumstances is likely to cause significant damage. This study aimed to detect respiratory diseases in livestock at an early stage. First, the temporal variation of heartbeat and respiration frequencies was analyzed considering the body-conducted sound of pigs in the time-frequency domain. Second, the effectiveness of time-frequency parameters was validated by training and adapting a machine-learning model to detect the disease. The proposed neural network, which has time-frequency parameters as the input, was found to require a shorter training time and exhibited higher accuracy in constructing the disease detection model. These results are expected to aid in automating livestock health management.