Abstract: Cough is a very common symptom in respiratory diseases. Objective analysis on the frequency and intensity of cough by pattern recognition algorithm can provide more valuable clinical information for patients with chronic cough and help them with cough tracking and diagnosis. Cough detection is the basis of cough diagnosis and analysis in clinical continuous recordings. In this article, we consider cough detection problem as a binary classification ones and make use of classifier to segregate cough event from background noise for the purpose of cough detection. We propose a novel high-frequency subband features method on the basis of in-depth study of the spectral distribution of cough .It is found that the energy of cough signal is distributed widely and concentrated in the high frequency region, which is very different from spectral patterns of speech signals. So in experiment, we firstly extract subband features whose frequency region varies from low frequency to high frequency using filter banks then find the performance of high frequency-subband features is superior to that of low frequency-subband.Finally Thehigh-frequency subbandmethod use high-frequency filter to get corresponding high frequency signal before extracting the features of cough. The method synthesizes the experimental data under the condition of different noisy type and SNR(signal to noise ratio) then compares and analyses the performance of different feature extraction methods under specific noisy conditions. Experimental results demonstrate that the method based on high-frequency subband features achieves substantial performance improvement in recognition compared with traditional audio feature extraction method.