In 12 patients, a SEEG working-memory error predictor retained modest accuracy between nearby trials (AUC 0.566)

Errors cluster in time in ds006136. Thus a SEEG error predictor can get its accuracy from slow state drift and not from the trial. We tested this with the archived band-power features of sci da6a9c25 and a replication of its leave-one-subject-out decoder (L2 logistic, C = 0.1, encoding + delay power, 12 subjects). The plan was written before the scores. The global AUC replicated (median 0.590, 10/12 above 0.5). Primary test: AUC only over error/correct pairs 1-5 trials apart. Median 0.566 (95% CI 0.524-0.626), 10/12 above 0.5, one-sided exact Wilcoxon p = 0.0061. Local minus global AUC: median -0.014 (CI -0.026 to -0.001), p = 0.077. A detrended score gave 0.571 (p = 0.0034). The trial score had a slow component (lag-1 autocorrelation 0.18, 12/12 positive). A behaviour-only predictor (error rate of the previous 10 trials) gave 0.542 and was not significantly different from the neural predictor (p = 0.20). In one subject (P7), the AUC decreased from 0.64 to 0.54 between neighbouring trials. Conclusion: most of the error information is trial-level, which a closed-loop trigger needs. Limits: offline, n = 12, the trial index is assumed to be chronological, z-scoring uses all session trials, AUC near 0.57 is too small for reliable single-trial triggers alone.

stereo-EEG · working memory · error prediction · trial-level decoding · area under the curve · state drift · closed-loop brain-computer interface