Conventional brain tumor diagnosis waits for a mass to grow large enough to show up on an MRI. By that point, the damage is already done — treatment windows are narrower, and survival outcomes are far worse than they needed to be. This work takes a different angle. Rather than searching for a physical lump, we ask a more fundamental question: what does a growing tumor do to the brain's electrical behavior, and can an AI system learn to recognize that disruption early? Our method is built around Phase-Amplitude Coupling, or PAC — a measurable property of how high- and low-frequency brain waves coordinate with each other. Tumor-induced inflammation disrupts this coordination well before anything appears on structural imaging. We trained a Transformer-based deep learning model on synthetic EEG signals from 900 virtual subjects, paired with three behavioral biomarker features: eye-tracking variance, gait asymmetry, and speech syntax irregularity. On the held-out test set, the model reached 87.4% classification accuracy and an AUC-ROC of 0.93 for the highest-risk class. Attention analysis showed the model learned to focus on theta-gamma PAC features across frontal and temporal channels — precisely where neuroscience literature would predict disruption to appear. The results demonstrate that this detection framework is computationally viable, even if clinical validation lies ahead.
Kudale et al. (Mon,) studied this question.