Using Computer Intelligence for Depression Diagnosis and Crowdsourcing

Humans are reasonably good at sensing when others are depressed. In particular, our hearing is attuned to the voice-related changes often exhibited during depression. But could computers also be taught to “hear” depression in everyday conversations outside a lab, even to the point of being able to make a clinical diagnosis? In their article “Self-Reported Symptoms of Depression and PTSD Are Associated with Reduced Vowel Space in Screening Interviews,” Stefan Scherer and his colleagues investigate such an automated approach to diagnosis (IEEE Trans. A ective Computing, vol. 7, no. 1, 2016, pp. 59–73). They focus on the acoustic changes in vowel sounds— particularly on vowel reduction— commonly observed among individuals with psychological disorders such as depression or neurological disorders such as Parkinson’s disease. To make diagnosis less reliant on clinicians’ or caretakers’ subjective impressions, the authors use SimSensei, a virtual human interviewer that “listens” to patients’ voices during unprompted conversations. They argue that this system could lead to more natural patient behavior during interviews, more objective evaluations, and reduced costs. Scherer and colleagues suggest that employing virtual agents o ers clinicians increased control over the choice of stimuli. Because virtual humans can be programmed a priori, they avoid the predispositions or “impurities” that human partners might introduce. The authors also propose that virtual human interviewers can reduce patients’ stress, fear, and sense of being judged. Ultimately, this encourages patients to speak more freely, thus providing richer data for analysis. In the study, the SimSensei virtual human system interviewed both US military veterans and nonmilitary individuals to identify—based on their speech characteristics—those at risk for posttraumatic stress disorder (PTSD) or major depression. All had been “coded for depression and PTSD” by standard psychiatric questionnaires. The findings support the assumption that vowel reduction was indeed related to depression or PTSD diagnosis.