Scientists have created an artificial neural network that can determine whether a person has signs of depression from a person's speech, recorded in audio or text format. At the same time, the operation of the algorithm does not depend on the context of the conversation, that is, it does not matter what the person is talking about, the neural network can find alarming signals even in the most abstract conversation. The research results will be presented at the Interspeech 2018 conference; its results can be read on the website of the Massachusetts Institute of Technology.
To date, there are already algorithms that can track the patient's answers to the doctor's questions and make a diagnosis based on them. Such neural networks analyze what the patient has said, and on the basis of this they decide whether a person has depression or not. Typically, doctors ask about previous mental illness, lifestyle, and so on. As the authors of the new study note, such conversations bear little resemblance to ordinary conversations that a person leads in life. Therefore, their goal was to teach a neural network to analyze not what a person says, but how he does it.
To train the neural network, the authors used more than 140 audio, video, and text files with recordings of interviews with patients with various mental disorders. First, specially invited experts manually rated each interview on a scale from 0 to 27. If the doctor gave the interview a “score” higher than 15, then the patient is showing signs of depression. After training, the neural network began to set points itself. The accuracy of the diagnosis (it was assessed in comparison with the verdict of the experts) averaged 77%.
The created algorithm analyzes the patient's speech, while the topic of conversation can be the most abstract. According to the authors, this technology in the future can be very useful for people who cannot see a doctor or do not want to. On the basis of the new development, it is possible to create, for example, a mobile application that will track a person's messages and his phone calls and only based on this information detect signs of depression.