Project: Temporal and Multi-modal Gaze Dynamics as Markers of Anxiety Severity in Webcam-Based Screening
Description
Extracting meaningful patterns from several synchronized time series at once, rather than from one signal in isolation, is a hard and relatively under-explored problem in visual analytics: the informative structure often lies not in any single channel's shape, but in how multiple channels move together or apart over time, which standard per-channel aggregation cannot capture. Webcam-based mental health screening platforms such as Anima (the company we are working with), are a concrete instance of this: gaze, head pose, and facial-landmark signals are recorded together throughout each session, yet are reduced to a single score derived from gaze alone, that is meant to reflect anxiety severity. This discards temporal order and any cross-signal relationship that might carry further information about that severity. A second, coupled problem is interpretability: once such multivariate, temporally-structured patterns are found, explaining which signals and time-windows drive a result has no established convention, since existing explanation methods are built for flat, non-temporal feature sets. Finding this structure, testing whether it relates to anxiety severity beyond the current score, and making it legible to a target Online user are the three problems this thesis addresses.
Details
- Student
-
DRDaniel Ris
- Supervisor
-
Fernando Paulovich