Personalised Deep Learning for Monitoring Depressed Mood from Speech

We utilise a longitudinal dataset of 17 526 speech samples collected from 30 patients with major depressive disorder and 11 sub-clinically depressed individuals to perform a personalised prediction of depressed mood. The data has been recorded via a smartphone app over a two-week ecological momentary assessment with three recording sessions per day. Each session's speech samples are accompanied by a self-assessed rating on the discrete visual analogue mood scale (VAMS) from 0-10. As these ratings are highly subjective, a personalised machine learning method is leveraged. For this purpose, the beginning of the recording period is utilised to train both a shared model backbone, and adapt personalised layers added at the end to each speaker's speech. Our approach yields a Spearman's correlation coefficient (p) of 0.79 on the test set, compared to the non-personalised baseline of p=0.61. Furthermore, we analyse our results with regard to the type of speech sample - reading three depression-related questions, answering them, and freely formulating an uplifting spontaneous thought. Here, we find that personalisation boosts performance across all types, especially for the fixed content question readings. Overall, our work highlights the efficacy of personalised machine learning for depressed mood monitoring.