A machine-learning framework for inferring latent mental states from cursor and touchscreen activity.
Scalable mental-health assessment remains a major bottleneck for accessible and equitable care. MAILA asks whether everyday human-computer interactions, especially cursor and touchscreen activity, contain reliable information about a person’s current mental state. In this work, we show that digital behavior can be used to infer clinically relevant dimensions of distress and wellbeing with surprisingly high accuracy.
MAILA, short for MAchine learning framework for Inferring Latent mental states from digital Activity, was trained on 18,200 recordings from 9,500 participants and 1.3 million self-reports. The broader goal is to establish human-computer interaction as a non-verbal, scalable modality for digital phenotyping.
Cursor and touchscreen patterns predicted 13 clinically relevant dimensions of mental health, linking routine digital activity to meaningful variation in distress and wellbeing.
Frozen MAILA models generalized across tasks, devices, and independent datasets. MAILA also tracked systematic changes associated with time of day, arousal, and valence, suggesting that it captures dynamic state fluctuations rather than only static traits.
MAILA recovered structure that standard questionnaires only partly reflect and improved the decoding of belief instability when combined with self-report. This suggests that human-computer interactions contain non-verbal information about mental health and cognition that is not exhausted by language.
The project combines:
MAILA suggests that passive digital behavior can become a new measurement modality in psychiatry: scalable, low-cost, and less dependent on language. If validated further, this approach could support screening, longitudinal monitoring, and more precise models of mental-health change.