Minimal Reservoir Computing Generates Hippocampal-Inspired Stimulation for Open-Loop Modulation of Cortical Networks In Vitro.
BACKGROUND High-frequency deep-brain stimulation can reduce seizure burden but imposes a large stimulation load on the tissue, and its effects depend on both frequency and temporal patterning. We tested whether a compact reservoir model could generate a hippocampal-inspired event schedule and whether that schedule could modulate neuronal activity in vitro with fewer nominal pulses than 50-Hz stimulation. METHODS A 10-unit leaky reservoir with a one-dimensional affine ridge readout transformed cortical input recordings into CA3-inspired outputs. Retrospective computational validation used paired cortical-CA3 recordings, training-only scaling, removal of at least the first 10 s of every generated trace, date-grouped held-out evaluation, multiple spectral and temporal metrics, objective-function ablation and surrogate signals. One offline biomimetic schedule was then delivered open loop to primary rat cortical cultures on multielectrode arrays and compared with periodic 0.16-Hz and 50-Hz stimulation. RESULTS 17 of 19 paired recordings passed the pre-specified transient diagnostic. In 11 scorable date-grouped held-out recordings, the reservoir had modestly lower median log-PSD RMSE and autocorrelation error than linear and lagged-ridge baselines. When applied to primary cortical cultures, the primary pooled analysis of all 11 biomimetic-stimulated MEAs yielded a mean firing-rate slope of 0.7655 relative to the null value of 1 (95% CI for the mean slope, 0.491-1.040; p = 0.0865). In separate MEA-level fold-change analyses of all 11 biomimetic-stimulated MEAs, normalized firing rate was 0.8458 and normalized burst rate was 0.7792, both significantly reduced relative to baseline. A post hoc exploratory classification identified reduced firing-rate slopes in 4 of 11 MEAs. CONCLUSION These results provide an open-loop in vitro proof of concept that a low-event, irregular stimulation schedule can be associated with heterogeneous modulation of cortical-network activity. Periodic stimulation at the same nominal mean event rate did not produce a statistically significant modulation under the conditions tested. The irregular schedule may contain timing-related properties that contribute to the observed response, but the specific contribution of ESN-derived temporal ordering remains unresolved. Future work should test appropriate temporal controls and evaluate the approach in closed-loop and epilepsy-model experiments.