When Gameplay Becomes Stressful: A Preliminary Exploration of Sensor-Free Stress Modelling in Virtual Reality Shooter Games

Can a model trained only on low-pressure gameplay detect behavioural changes associated with stress without physiological sensors or stress labels during training? We investigate this question in a controlled virtual reality (VR) first-person shooter using an LSTM-based variational autoencoder (LSTM-VAE). The model was trained on tutorial sequences and evaluated during a subsequent survival phase using gaze, head and controller kinematics, or button inputs. Reconstruction error was converted into a continuous baseline-deviation score and compared with retrospective DANTE stress annotations from N=38 participants. The strongest condition, HMD+MC/Single, showed a statistically reliable but small association with self-reported stress (r=.214, approximately 4.6% shared variance), whereas binary separability remained close to chance. These findings do not support a deployable stress detector or establish that the recurrent variational architecture is superior to simpler alternatives. They instead provide preliminary evidence that native VR telemetry contains a small, consistent stress-related signal. Predictive processing and free energy motivate the baseline-deviation formulation, while reconstruction error is the operational quantity evaluated in the experiments.