Cross-patient SEEG band-power scores increase with the number of working-memory items lost

In 13 sessions from 13 patients performing a load-3 object working-memory task, a leave-one-subject-out shrinkage LDA trained to distinguish perfect from imperfect trials produced scores that were graded with the number of items lost. For held-out trials, scores were higher for 2–3 items lost than for 1 item lost (mean AUC 0.635, 95% CI 0.580–0.691; 9/9 eligible subjects above 0.5; exact two-sided sign-flip p = 0.0039). Scores were also higher for 1 item lost than for perfect trials (AUC 0.565, 95% CI 0.526–0.600; 10/12 subjects; p = 0.011), and for 2–3 items lost than for perfect trials (AUC 0.703, 95% CI 0.644–0.766; 9/9; p = 0.0039). The mean Spearman correlation between score and items lost was 0.154 (95% CI 0.095–0.213; 12/13 positive; p = 0.0007). A sensitivity decoder trained only on perfect and 1-item-loss trials gave the same ordering (AUC 0.641, 95% CI 0.585–0.697; 9/9; p = 0.0039). These results support a graded score in this dataset, strongest for larger losses. Only nine subjects had at least five trials with 2–3 items lost. Channel-mean features, lack of item identity and response order, possible attention or arousal effects, and use of all trials for held-out-session z-scoring limit interpretation and causal use.

working memory · SEEG · brain-computer interface · shrinkage LDA · graded error score · band power · leave-one-subject-out