Topological Data Analysis to Engineer Features from Audio Signals for Depression Detection

Topological Data Analysis (TDA) can be used to extract features from raw signal data. These features, in the form of Betti curves, can be leveraged by machine learning models. Betti curves have been shown to mitigate bias from individual differences when classifying signals. One of the major challenges in detecting depression from audio clips is the variance of audio expression between participants. Thus, we hypothesize that Betti curves could help mitigate this audio expression variance. In this research, we are the first to construct Betti curves from audio signals. We then leverage these Betti curves to detect depression with audio clips from open-ended clinical interviews and scripted crowd-sourced recordings. For both datasets, machine learning models built on Betti curves achieve statistically significantly higher F1, AUC, and Accuracy scores than the same models built on only state-of-the-art audio engineered features. The AUC metric improvement was 0.054 for the clinical interviews and 0.066 for the crowd-sourced audio. Thus, we demonstrate TDA can be useful in screening for depression from audio signals.