The aim of this thesis work was to design and develop a fully adaptive feedback loop system for virtual reality that uses real-time physiological signals to automatically modulate game experiences according to predicted user stress levels. The system comprises a context-independent plugin that allows a game application to automatically adapt in-game elements, balancing the user's state and maintaining a designer-specified "target stress level". It uses a local server for data processing and real-time stress assessment, taking GSR biosignal and producing a binary prediction value ("non-stressed" or "stressed"); the inference result is then streamed to all connected VR applications. As part of the work, the deep neural network for automatic stress recognition was designed, trained, and tested.