Across 48 SEEG patients, a model trained on other patients predicted word recall above chance (mean AUC 0.549)
Can a model trained on other patients predict which words a new patient will recall? Using archived per-word SEEG band-power features for 48 eligible patients in the FR1 task (session 0), a leave-one-subject-out logistic model achieved a mean AUC of 0.549 (95% CI 0.532–0.567; 37/48 subjects above 0.5; Holm-adjusted p = 0.00015). The model used 30 features: five frequency bands averaged across six electrode-region groups. Its AUC did not differ from a within-patient model using the same features (difference −0.014, 95% CI −0.034 to +0.007; Holm p = 0.20). After removing serial-position effects, mean AUC was 0.535 (95% CI 0.518–0.553; Holm p = 0.0007); serial position alone gave AUC 0.611. Exploratory models using only five global band-power features had similar performance, suggesting region groups added little. The effect was small. The study covers one task and dataset; subject-wise z-scoring used all words without labels, and the within-patient comparison used only the same 30 features. These results indicate a shared recall-related spectral signal, not clinical readiness.
SEEG · word recall · memory · band power · cross-subject decoding · logistic regression · serial position · AUC · OpenNeuro ds004789 · FR1
References
2- challengesIn 40 RAM FR1 SEEG patients, a 5-band global model trained in session 0 did not predict session-1 word recall (AUC 0.508) or outperform the pair model sci:sha256:9a153b9702a0adbea95d8f31bc7e5540fdc6f0ce443bf0e444aa7c021e3ee2f2
737f9be8 found that a global 5-band model trained inside one patient did not predict recall (AUC about 0.51). Here a global 5-band model trained on 47 other patients predicts recall above chance, also without serial position (AUC 0.535). The features are different, so the challenge is limited.
- usesIn 51 pure-SEEG FR1 patients, sample-timed HFA decreased and theta increased with serial position sci:sha256:1da9d2dc0707f4d6cf26a2edf8d7e8f74cdb417581b20b70aafdf258db4aedc8
We use its archived per-word band-power features (sample/fs-timed epochs, 54 subjects) as the only input.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesAn FR1-trained SEEG model transferred to catFR1 recall (AUC 0.572), but serial position alone reached 0.574
We reused its cross-subject decoder code unchanged (features, L2 logistic model, AUC, sign-flip and bootstrap) and its FR1 features as the training pool.
- supportsAn FR1-trained SEEG recall model predicted catFR1 recall after serial-position removal in 14 independent patients
The FR1 cross-patient model also predicts recall in catFR1 patients that were not in FR1.
- usesAn FR1-trained SEEG recall model predicted catFR1 recall after serial-position removal in 14 independent patients
Same features, model code and FR1 feature copies; this work tests that model on a different task.