Automating anxiety detection using respiratory signal analysis

In this paper, we explored the development of an anxiety detection (AnD) system using the respiratory signal as its input. Time and frequency domain statistical features derived from breath-to-breath (BB) interval series of respiratory signal is input to a support vector machine (SVM) backend classifier. We used data from normative population, individuals with anxiety disorders and regular meditators for validating the effectiveness of the system. We experimented with different kernels for the backend SVM classifier in our baseline system, and note that the best results were obtained for the polynomial kernel, a classification accuracy of 69.23%. It may be noted that for classification using SVM, either the features should match the kernel, or the kernel that matches the features should be selected for optimum performance. Often, this process is very difficult, owing to the difficulty in identifying a matching kernel. Alternatively, we may transform the feature vectors to a higher dimensional linear space, and then use SVM with a linear kernel. We used Fisher vector encoding (FVE) for mapping the features to a higher dimensional linear space. Also, we examined principal component analysis (PCA) and covariance normalization on the input features and the transformed feature vectors, in an effort to reduce the effect of patient specific variations in the signal to improve the performance. A performance improvement of 7.69% absolute using PCA-AnD, 15.38% absolute using FVE-AnD, and 15.38% absolute using CVN-AnD, over the baseline system were obtained. Further, we combined FVE and CVN, and obtained FVE-CVN system with a classification accuracy of 92.30% absolute, which is 23.08% absolute improved over the baseline system.