Raw LDA probabilities were overconfident in seven SEEG patients, while Platt calibration did not significantly lower ECE

We tested whether shrinkage linear discriminant analysis (LDA) probabilities were overconfident on later trials and whether a Platt map fit on training data only reduced expected calibration error (ECE). This replication used archived 70–150 Hz high-gamma activity (HGA) window means from seven patients in OpenNeuro ds006253: rest from −0.5 to 0 s before the stimulus and movement from 0 to 0.5 s after mouse-movement onset. For each patient, we trained on the first 60% of trials in stimulus-onset order and tested on the last 40%; the Platt map was fit using five-fold out-of-fold training scores. Raw probabilities were overconfident in all seven patients (mean confidence minus accuracy +0.115; one-sided exact sign-flip p = 0.0078). The primary endpoint, ECE(Platt) − ECE(raw), averaged −0.047 and was negative in five of seven patients, but did not meet the 0.05 threshold (p = 0.055). Secondary, uncorrected differences favored Platt calibration for log loss (−0.112, six of seven, p = 0.016) and Brier score (−0.016, six of seven, p = 0.023). Mean overconfidence after Platt calibration was +0.030; two patients became underconfident. The overconfidence finding replicated, while the primary ECE benefit was only partly replicated. This was an offline analysis of one session per patient (n = 7), with balanced fixed windows and HGA only; movement windows followed the stimulus, so classes also differed in stimulus state.

probability calibration · linear discriminant analysis · stereo-EEG · brain-computer interface · expected calibration error · high-gamma activity · Platt scaling · mouse movement