Neural, Behavioral, and Clinical Outcomes of a Voice-based AI Coach for Depression and Anxiety: A Phase 2 Randomized Trial

Importance: Artificial Intelligence (AI) voice-based chatbots may improve access to evidence-based psychotherapy for depression and anxiety, but their therapeutic mechanisms and outcomes are largely unknown. Objective: To investigate the mechanisms and potential efficacy of Lumen, an AI voice-based coach delivering problem-solving treatment (PST), among adults with untreated, clinically significant depression and/or anxiety. Design: Phase 2, 3-arm randomized clinical trial. Setting: A public university and affiliated medical center in Chicago, Illinois. Participants: Adults with a Patient Health Questionnaire-9 score of 10-19 and/or a Generalized Anxiety Disorder Scale score of 10-14 were randomized to Lumen-coached PST (n=100), human-coached PST (n=50), or waitlist control (n=50). Interventions: PST was delivered in 4 weekly sessions followed by 4 biweekly sessions, by Lumen - a custom-developed, rule-based AI application on Amazon's Alexa platform - or by a human coach via videoconferencing. Main Outcome(s) and Measure(s): The primary mechanistic outcome was change in right dorsolateral prefrontal cortex (dlPFC) activation for cognitive control, assessed with functional neuroimaging. Secondary outcomes were patient-reported, theory-based behavioral targets (e.g., problem-solving) and clinical symptoms (e.g., psychological distress). Results: Among 200 participants (mean (SD) age, 36.6 (11.9) years; 77% women; 25% Black American; 28% Latino), change from baseline in right dlPFC activity at 18 weeks did not differ significantly by treatment arm. Compared with waitlist control, Lumen-coached participants had significantly greater improvements from baseline to 18 weeks in overall problem-solving ability (mean difference, 1.04; 95% CI, 0.23 to 1.84]) and psychological distress (mean difference, -3.56; 95% CI -5.69 to -1.43) due to depressive and anxiety symptoms. Lumen- and human-coached PST did not differ significantly on either measure. One serious adverse event (hospitalization), unrelated to this study, was reported. Conclusions and Relevance: In this phase 2 experimental-therapeutics RCT, neither AI- nor human-coached PST significantly engaged the prespecified neural target for cognitive control. However, both interventions improved patient-reported problem-solving and psychological distress relative to waitlist control, providing exploratory evidence of a non-neural candidate mechanism and a clinical signal warranting future testing. Trial Registration: clinicaltrials.gov, NCT05603923