Phase 2 RCT finds AI-coached therapy improves depression and anxiety symptoms versus waitlist

Importance: Artificial intelligence (AI) voice applications may address unmet depression and anxiety treatment needs. However, their therapeutic utility is largely unknown. Objective: To investigate the mechanisms and efficacy of a voice-based AI coach, Lumen, delivering problem-solving treatment (PST) for depression and anxiety. Design: Phase 2, 3-arm randomized trial. Setting: A public university and affiliated medical center in Chicago, IL. Participants: Adults with clinically significant depression and/or anxiety were randomized to Lumen-coached PST (n=100), human-coached PST (n=50), or waitlist control (n=50) for 18 weeks. Interventions: PST was delivered by Lumen on Amazon's Alexa platform or a human coach via videoconferencing in 4-weekly and then 4-biweekly sessions. Main Outcome(s) and Measure(s): Change in activation of the right dorsolateral prefrontal cortex (R dlPFC) for cognitive control using functional neuroimaging was the primary mechanistic target measure. Symptom outcomes were assessed using Hospital Anxiety and Depression Scale. Other secondary measures included left dlPFC and bilateral amygdala activation and patient-reported problem-solving abilities, dysfunctional attitudes, worry, positive and negative affect, disability, and work productivity. Statistical analyses used t-tests and ordinary least square regression. Results: Participants were on average 36.6 years old (SD=11.9), 77% women, 25% Black, 29% Latino, and 21% Asian. At 18 weeks, change from baseline in R dlPFC activity did not differ significantly by treatment arm. Compared with waitlist control, Lumen-coached PST resulted in significantly greater improvements in symptoms of psychological distress (between-group mean difference=-3.56, 95%CI [-5.69, -1.43]), depression (-1.15, 95%CI [-2.25, -0.04]), and anxiety (-2.43, 95%CI [-3.72, -1.14]). Similarly, Lumen-coached participants had significantly greater improvements in problem-solving abilities, worry, positive and negative affect, disability, and work productivity, but not in the secondary neural targets or dysfunctional attitudes, compared with waitlist control. Lumen- and human-coached PST did not differ significantly for any of the measures; additionally, improved problem-solving abilities correlated with improved symptoms for both modalities. One serious adverse event involving hospitalization, unrelated to the study, was detected. Conclusions: Although differences in neural target changes were null, a voice-based AI coach improved measures of symptoms, behavior, affect, and functioning. Although Lumen showed promise compared to human delivery, confirmatory noninferiority testing is warranted. Trial Registration: NCT05603923