A training mean gave less error than two fixed models of processed SEEG patterns in forward tests of all 18 tested MetaRDK patients

We tested processed high-gamma SEEG patterns from 18 MetaRDK patients. The fixed analysis used 1,658 channels and 1,415 later test trials. Adding previous-trial channel history reduced mean patient error by 1.38% relative to a behavior model. Its 95% patient interval was 0.42% to 2.34%. A later reference check changed the interpretation. Both fixed models had more error than the training mean in every patient. Their mean error increases were 58.45% and 56.57%. An added history-only model reduced mean error by 0.68% over the training mean. Its interval was minus 0.22% to 1.63% and included zero. These results do not establish reliable signal prediction from history. They show why model comparisons need a simple reference. The package preserves code, compact data, source versions, exact retrieval ranges, plans, model outputs, and figures. Fresh checks verified all 3,710 retained source windows. Offline results repeated exactly. The tested scope is one task and later trials within a reused clinical subset.

stereoelectroencephalography · SEEG · high-gamma activity · trial history · neural signal prediction · ridge regression · forward testing · MetaRDK