Psychiatric care has historically relied on episodic clinical encounters. A 2025 narrative review published in the Journal of Multidisciplinary Healthcare (Narimani et al.) identifies a shift with AI systems now enabling continuous monitoring by integrating data from wearables, smartphones, and acoustic inputs. One cited study by Lee et al. demonstrated that AI models predicted depressive episodes with 91% accuracy using wearable devices and mobile applications, though the authors note this figure derives from a single-site cohort and should be interpreted with caution.
The approach underpinning much of this work is digital phenotyping, the passive collection of behavioural signals such as screen usage, geolocation patterns, and sleep-wake rhythms. A scoping review published in JMIR (August 2025) examined 42 peer-reviewed studies using machine learning on wearable and smartphone data in clinically diagnosed populations, finding consistent associations between passive behavioural features and mood disorder indicators.
Large language models are now being applied directly to this sensor data. A September 2025 paper in BMJ Mental Health evaluated whether LLMs could interpret psychiatric digital phenotyping outputs, finding they performed comparably to human clinical staff on pattern identification tasks, whilst noting hallucinations and sensitivity to low data quality as current limitations.
The first randomised controlled trial of a fully generative AI chatbot, Therabot, developed at Dartmouth, published results in NEJM AI (2025) showing significant symptom reductions across major depressive disorder and generalised anxiety disorder. It is important however to also consider opposite tendencies such as the one found in a 2026 systematic review in MDPI Electronics covering 205 studies. The research concludes that despite promising performance, most evaluations rely on small, non-longitudinal datasets, and standardised clinical validation remains absent.