Joint learning for voice based disease detection

Abstract Voice analysis provides a non-invasive way for disease detection, in which most methods only consider a single audio, although different audios contain complementary information and a fusion of them is beneficial. In this paper, a novel model JOLL4R (JOint Learning based on Label Relaxed low-Rank Ridge Regression) is proposed to fuse audios for voice based disease detection. First, the model couples the regression losses from two audios together to jointly learn a transformation matrix for each audio. Secondly, the conventional zero-one regression targets are relaxed by the ϵ-dragging technique so that the margins between different classes are enlarged. Third, low-rank constraint is imposed to exploit the correlation structure among different classes. The proposed algorithm not only enables to consider multiple audios, but also adjusts the weight of each audio adaptively. Due to the design of losses coupling, ϵ-dragging technique, and low rank constraint, high performance is achieved. Experiments conducted on two disease detection tasks, each with six types of fusion, show that our fusion approach outperforms the case of using a single audio and another two fusion methods. Finally, key factors in JOLL4R are analyzed.