Detection of Universal Cross-Cultural Depression Indicators from the Physiological Signals of Observers

We conducted a pilot study experimenting with neural network techniques to use the physiological signals of untrained observers to classify the depression levels of variously depressed people in videos speaking a language the observers did not understand. As the dataset was highly imbalanced, noisy and thus extremely sensitive to relative class sizes, we developed a technique for dynamically oversampling the smaller classes both prior to and during training to approximately align training prediction rates for each class with knowledge of the prevalence of different levels of depression. In predicting the depression levels to a final accuracy of 57.9% over four classes and 78.9% over three classes we demonstrate the likelihood that universal cross-cultural indicators of depression exist. In addition, that some people's automatic physiological responses to these indicators are strong enough that they can be used to predict depression categories of people to a significant degree of accuracy even when the observer does not understand the language the person is speaking. The final accuracy rate is significantly better than the diagnosis rates of doctors speaking to patients in their own language. The results show the potential these techniques have to improve diagnosis of depression, especially in areas with limited access to mental health professionals. This innovative approach demonstrates the importance of further experimentation in this area and research into universal cross-cultural depression indicators.