Benchmarking single-subject detectability of cognitive event-related potentials in healthy controls and minimally conscious state patients

Event-related potentials (ERP) may reveal residual cognitive processing in disorders of consciousness, but clinical translation is limited because group-level effects do not guarantee reliable single-subject detection. Few ERP paradigms have been benchmarked in healthy participants and compared directly with patients in disorders of consciousness. We tested a hierarchical ERP battery in healthy controls and patients in minimally conscious state (MCS) to (1) estimate upper-bound single-subject detectability in healthy participants, (2) compare detection rates with those of MCS patients, and (3) evaluate prognostic value for recovery of consciousness. Forty-eight healthy controls and 26 patients in MCS completed six auditory paradigms: multifeature mismatch negativity (MMN), passive and active oddball P300, the subject's own name, semantic oddball, and sentence N400. Effects were classified using blinded expert visual scoring, region-of-interest ANOVA, cluster-based permutation tests, and support vector machine decoding. Associations between binary ERP classifications and recovery were tested in 22 surviving patients at follow-up. In healthy controls, visual scoring identified MMN to at least one deviant and P300 in both oddball tasks in all 48 participants. Command following as demonstrated by task-related P300 enhancement between active and passive oddball conditions was detected in 81%. Command following was detected in 92% in the semantic oddball task. Sentence-congruity effects (N400 and/or P600) were identified in 69%, despite a strong group-level response, while the own-name effect was detected in only 35%. In patients, detection rates were lower: MMN was found in 27% for the short-duration deviant and 65% when any of five deviant types counted as positive; passive oddball P300 in 35% and active oddball P300 in 31%; sentence-congruity effects in 12%; active-versus-passive oddball modulation in 12%; and semantic oddball effects in 8%. Visual scoring and machine learning generally detected more effects than ANOVA and cluster-based testing, and agreement varied substantially across paradigms. Binary ERP classifications were not significantly associated with recovery (14 recovered, eight did not), although prognostic analyses were limited by sample size. Group-level ERP effects overstated single-subject detectability in healthy controls, whereas the absence of group-level effects in MCS obscured reliable ERP responses resembling those of healthy controls in a few individual patients. In MCS, ERP effects associated with semantic processing and command following were generally uncommon, and binary ERP indicators did not predict recovery in this sample. Our findings establish a benchmark for selecting and refining ERP paradigms and detection methods to assess residual cognition after severe brain injury.