Artificial Intelligence for Management of Major Depression: Initial Design, Progress, and Research Plans.

Background: Herein, we report on the initial development, progress, and future plans for an autonomous artificial intelligence (AI) system designed to manage major depressive disorder (MDD). The system is a web-based, patient-facing conversational AI that collects medical history, provides presumed diagnosis, recommends treatment, and coordinates care for patients with MDD. Methods: The system includes seven components, five of which are complete and two are in development. The first component is the AI's knowledgebase, which was constructed using Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression to analyze extensive patient medical histories and identify factors influencing response to antidepressants. The second component is a series of adjustments to the knowledgebase designed to correct algorithm bias in patient subgroups. The third component is a conversational Large Language Model (LLM) that efficiently gathers patients' medical histories. The fourth component is a dialogue management system that minimizes digressions in the LLM conversations, using a topic network statistically derived from the AI's own knowledgebase. The fifth component is planned to enable real-time, human-in-the-loop monitoring. The sixth component is an existing analytical, non-generative module that provides and explains treatment advice. The seventh component is planned to coordinate care with clinicians via closed-loop referrals. Results: In component 1, the AI's knowledgebase correctly predicted 69.2% to 78.5% of the variation in response to 15 oral antidepressants. Patients treated by AI-concordant clinicians were 17.5% more likely to benefit from their treatment than patients of AI-discordant clinicians. In component 2, the use of the system required adjustments to improve accuracy for predicting the responses of African Americans to four antidepressants and no adjustments were required for the remaining 10 antidepressants. In component 3, the conversational intake efficiently covered 1499 relevant medical history events (including 700 diagnoses, 550 medications, 151 procedures, and 98 prior antidepressant responses). In the fourth component, the dialogue management system was effective in maintaining a long dialogue with many turns in the conversation. In the sixth component, the advice system was able to rely exclusively on pre-set text. An online ad campaign attracted 1536 residents of Virginia to use the advice system. Initially, a focus group of clinicians was skeptical of the value of the advice system and requested more prospective studies before they would implement the system in their clinics. When the system was redesigned to advise patients at home, clinicians were willing to receive referrals from the system and discuss the advice of the system with their patients. Conclusions: Further research is needed to refine and evaluate the system. We outline our plans for a prospective randomized trial to assess the system's impact on prescription patterns and patient outcomes.