SEEG working-memory decoding improved with more calibration trials but had not plateaued by 12

We asked whether the number of labeled trials limits within-patient decoding of verbal working-memory load from stereo-electroencephalography (SEEG). We used archived maintenance-period (3–6 s) log band-power features from bipolar SEEG channels for 15 patients in OpenNeuro ds004752, comparing correct set-size 4 and 8 trials. A shrinkage linear discriminant decoder used either five frequency bands (ALL) or high gamma alone (HG). In 400 random draws per patient, each draw held out three trials per class and evaluated nested training sets. In the 12 patients with enough trials for the primary comparison, increasing ALL-decoder calibration from 6 to 12 trials increased held-out AUC by 0.054 (95% bootstrap CI 0.033–0.074; 10/12 patients positive; exact one-sided sign-flip p = 0.001). Across all 15 patients, the mean gain was 0.050 AUC per doubling of calibration trials (13/15 positive; p = 0.0001). Group mean AUC rose from 0.600 at 4 trials to 0.689 at 12 trials and did not flatten in the measured range. High-gamma-only performance showed a smaller gain per doubling (0.038); its difference from ALL was not statistically significant at 6 or 12 trials (p = 0.076 and 0.080). Two patients remained near chance for the ALL decoder across training sizes. A descriptive extrapolation beyond the observed range estimated about 25 trials for AUC 0.75 and 47 for AUC 0.80; these estimates were not tested. The analysis is limited to one dataset, one session, one feature set and one decoder, with only three test trials per class per draw, training sizes up to 12 trials in the primary analysis, and random trial order.

stereo-electroencephalography · working memory · calibration trials · learning curve · shrinkage LDA · AUC · brain-computer interface · OpenNeuro