Personalized Brain-Based Anesthetic-State Detection with Portable fNIRS and AI

Portable functional near-infrared spectroscopy (fNIRS) may support brain-based monitoring around local anesthesia, but between-patient variability and patient-overlapping validation can limit generalization. We examined affected-tooth classification before versus after anesthesia, exploratory Pre-session affected-versus-healthy classification using both resting-state and held-out healthy-tooth reference schemes, and validation inflation. We analyzed fNIRS data from 25 patients during percussion before ("Pre") and after ("Post") anesthesia. The primary cohort comprised 13 patients with complete pain abolition; 156 affected-tooth epochs were classified using Pre-session healthy-tooth responses as patient-specific references. Leave-one-subject-out (LOSO) validation tested each patient using models trained on the others. A random forest using a fixed, data-informed three-feature oxyhemoglobin (HbO) panel achieved mean area under the receiver operating characteristic curve (AUC) = 0.752 (95% confidence interval [CI], 0.624-0.865; one-sided exact conditional p = 0.0057). Fold-nested selection yielded AUC = 0.740. Referencing improved estimates for two of four classifiers. In Pre-session analyses, resting-state and split-half referencing increased affected-versus-healthy AUC from matched raw values of 0.513 and 0.524 to 0.731 (12 patients) and 0.782 (21 patients); both gains survived eight-comparison adjustment. Because tooth blocks followed a fixed order, these pain-related estimates may include block-position effects. Across 32 feature-based configurations, pooled-epoch cross-validation exceeded LOSO by a mean of 0.159 AUC (95% CI 0.078-0.252); inflation was also positive in all three aligned waveform models. These results support proof-of-concept held-out-patient classification of pain-related affected-tooth responses before versus after anesthesia, exploratory referenced affected-versus-healthy classification before anesthesia, and patient-disjoint validation for estimating performance in new patients.