Chatbots as Mental Health Assistants: The Promise and Perils of Digital Phenotyping: The Promise and Perils of Digital Phenotyping

Digital health technologies—such as smartphone apps and chatbots—have the potential to augment mental healthcare through more ecological and scalable assessments and interventions. In this presentation, we will review the current state of technology in digital mental health, with a case study of how an artificial-intelligence (AI) tool has been integrated into a recovery-center setting to monitor treatment progress and outcomes. First, we will provide an update on recent technological advances in “digital phenotyping”—using smartphone and computer devices to understand mental health through multimodal assessment (e.g., using screen, microphone, and camera inputs to analyze language, voice, and facial expressions) in a more ecological context (i.e., as part of someone’s lived experience and not just in the clinic). Second, we will discuss the most promising use cases for this technology, from self-screening and early detection of mental health problems to long-term monitoring of chronic psychiatric conditions, with a focus on how this technology can improve upon measurements collected from Likert-scale questionnaires like the PHQ-9. Third, we discuss the challenges and opportunities of implementing technological innovations into mental health care systems, using our firsthand experience of developing and deploying a digital tool for monitoring depression and suicide risk in the context of a recovery program. We have found that some people feel more comfortable conversing with a chatbot, which provides an opportunity to improve detection of distress for individuals who feel uncomfortable with face-to-face contact. We conclude that the potential benefits of these tools outweigh their risks, but their use should ideally be supervised by a mental health professional in the context of a broader treatment plan—both as a guard rail against “edge cases” (atypical inputs that cause an AI system to produce erroneous output) and to encourage user engagement and accountability.