Stimulated brain responses last twice as long as textbooks say: SafeLens builds its own null from resting data and tracks cortical evoked potentials to 1,040 ms
Synopsis
SafeLens pairs nonlinear machine learning with strict false-positive control for event-locked neural activity. It trains a random forest at each sliding-window position to separate event-locked trials from resting trials, then runs the same pipeline on set-aside resting recordings to build a model-matched empirical null, holding the family-wise error rate per channel. On simulated cortico-cortical evoked potentials injected into real resting intracranial EEG it resolved about 2.5 times as many polarity-resolved deflections as trial-average thresholding and the closest data-driven method, with 5 to 11 times fewer false components; in eight patients responses stayed above chance to 1,040 ms after stimulation, against 570 and 455 ms for the two benchmarks.
Interpretation
Event-locked neural responses can now be judged deflection by deflection, with presence, onset and offset boundaries and polarity, under a 5% family-wise error rate per channel, without a preset analysis window or an assumption that trials share a waveform. Most cortico-cortical evoked potential studies previously thresholded the trial average inside a window fixed in advance, while canonical response parameterization learned one shape and duration per stimulating and recording pair and let only amplitude vary across trials. On real resting intracranial EEG with simulated responses injected, SafeLens reported a response on 19 of 352 channels with nothing injected, 5.4%, within the range expected of a test at the nominal 5%; on 704 injected channels, false positives overlapping no true deflection accounted for 2.8%.
Across 616 gray-matter channels in eight patients with drug-resistant epilepsy, SafeLens called 67.0% of channels responsive, with a median response span of 337 ms, against 175 ms for canonical response parameterization. On the same trials, canonical response parameterization called 55.8% of channels responsive and the z-scored trial-average detector 48.9%; SafeLens found more responsive channels than canonical response parameterization in six of eight patients and fewer in none. With patients weighted equally, the share of channels significant stayed above a chance threshold to 1,040 ms for SafeLens, 455 ms for canonical response parameterization and 570 ms for the z-scored detector.
Perspective
Validation covers simulated responses injected into real resting intracranial EEG and single-pulse recordings from eight patients implanted with stereo-EEG; the method requires responses to end before the next event and enough set-aside resting data to build the null. The authors propose next steps including a larger and more heterogeneous patient cohort and other stimulation paradigms, simulations with more subjects' resting backgrounds, late components absent from some trials, deflections coinciding in time, realistic stimulation artifacts and fewer trials.
The authors note that deflection-level accuracy can be measured only on simulated data, that the injection was additive while real responses depend on tissue state, and that the simulation used resting recordings from only two subjects, so treating channels of one subject as independent may make p-values too small. The null relies on exchangeability between set-aside rest and inter-event rest, and resting data from a different state were not tested. The scale factor s can grow with a strong response and cost sensitivity, and the method needs a GPU to run efficiently and was tested only at 276 to 378 pulses per channel.