A Virtual Doctor for ADHD Based on Fully Automated EEG Reading

Recording a 128-channel EEG is routine, technician-administered clinical practice, available far beyond specialist centers. What is not routine is reading the recording: reliably identifying ADHD-consistent patterns in a face-processing EEG currently requires an experienced specialist able to look past ordinary group-mean statistics, and this expertise, not the acquisition step, is the actual bottleneck limiting how many patients can be diagnosed this way. This study is motivated by removing that bottleneck entirely: a self-contained "virtual doctor" that takes a standard, technician-acquired EEG recording and outputs a diagnosis-relevant classification on its own, with no expert reader in the loop, and without a human first deciding which EEG features are worth showing the model. That second dependency is the more fundamental obstacle: existing EEG classifiers are typically trained only on ERP components that already survived conventional group-mean statistics, so the algorithm inherits a human-curated feature set rather than reading the recording for itself. We recorded 128-channel EEG from 102 ADHD and 49 control participants during a face-matching task and built a five-modality pipeline, ERP, time-frequency (ITPC), functional connectivity (PLI), deep-learning classification (CNN and Transformer), and explainable AI (Grad-CAM, occlusion sensitivity, SHAP, and attention weights), in which the classifier itself, not a prior statistical screen, determines which signal is used. Conventional statistics identified only 4 of 8 predefined components as significant (P1, N1, N170, P300). Critically, an ablation experiment restricted to only the three components a human reader would discard (P2, N2, Central-Late) still achieved AUC = 0.733-0.782, close to the full-signal model, directly falsifying the assumption that statistical non-significance equals absence of diagnostic value, and demonstrating that an autonomous reader does not need a human expert to pre-select its biomarkers to reach diagnostically meaningful accuracy. Two independent explainability methods, Grad-CAM and Transformer attention, converged on the same 150-300 ms window as the decision-relevant signal, showing that the model's self-discovered features correspond to genuine, reproducible neurophysiology rather than noise, and reduced inter-trial phase consistency (ITPC) supplies a mechanistic account of why this window evades group-mean statistics. Together, these results establish the reading and decision core of a fully automated EEG interpretation system, one that keeps EEG acquisition exactly as it is performed today, by a technician, but removes the expert-reader bottleneck standing between that recording and a diagnosis.