Matlab version 2018a

All data were analyzed with Matlab version 2018a (Mathworks, RRID:SCR_001622) except the psychometric test results, the NASA task load index and the Karolinska sleepiness scale, which were analyzed with SPSS version 27 (IBM, RRID:SCR_002865). All outcome measures were tested for normality with the Kolmogorov–Smirnov (K–S) test, and subsequently characterized by their mean and standard deviation (SD). Comparisons of psychometric test results, normally distributed NASA task load index scales and both EEG metrics between the exposures were analyzed with a two‐tailed paired t‐test and reported as mean difference and 95% confidence intervals (CI). Comparisons of the non‐normally distributed NASA task load index and the Karolinska sleepiness scale between the exposures were analyzed with the nonparametric Wilcoxon signed‐rank test and reported as median and range. Descriptive statistics were generated to characterize the study participants. Differences were regarded as significant at p < 0.05, with the Bonferroni correction applied for multiple comparisons.

Participants were asked to stand still for 2 min to record brain activity during quiet stance. Subsequently, participants were familiarised with treadmill walking and the preferred walking speed was determined. During data recording, participants walked for 200 strides on an instrumented treadmill (M-Gait, Motek Medical) at their preferred walking speed (Fig. 1). Participants were asked not to talk and to limit head movements during the experiment to avoid the occurrence of artefacts in the electroencephalogram (EEG) recording.Fig. 1

The experimental setup during the measurements

Three-dimensional full-body movements were recorded using reflective markers placed on the body according to the Vicon Plug-in Gait Full Body model [24 ] using the Vicon Motion Systems (Oxford, UK, 100 Hz). In addition, the force plates embedded in the treadmill (Motek Medical, NL, 2000 Hz) were used for step detection. Electrical brain activity was measured using a high-density EEG with 126 Ag–AgCl electrodes (WaveGuard, ANT Neuro, The Netherlands) distributed according to the five percent electrode system [25 (link)]. EEG was recorded using a biosignal amplifier (REFA System, TMSi, The Netherlands, 2048 Hz) and MATLAB version 2018a (The MathWorks Inc. USA). Both recording systems were synchronised using a digital trigger.

The sequential forward selection (SFS) algorithm was applied to select significant features for model building. Before selection, all features were normalized by z-scores. Subsequently, the normalized features were forwarded into an iterative procedure of SFS. In the SFS, a feature that achieved the highest performance (i.e., accuracy) across the extracted features was selected first. Subsequently, from the remaining features, a feature that could further improve the performance in combination with the first selected feature was selected; the rest of the features were then selected as per this procedure until there was no further improvement in the performance. Based on the selected features, a logistic regression model based on a linear kernel was constructed. The feature selection and model construction were performed using the Statistics and Machine Learning Toolbox in MATLAB version 2018a (MathWorks, Natick, MA, USA).
To investigate if the proposed model could be used for tumor invasiveness prediction in the general setting, external validation was performed in an external validation cohort (n = 100) with lung cancer patients undergoing pGGN lung tumor resection.

All data were analyzed with Matlab version 2018a (Mathworks, Natick, MA, USA) except the psychometric test results, the NASA task load index and the Karolinska sleepiness scale, which were analyzed with SPSS version 27 (IBM, Armonk, NY, USA). All outcome measures were tested for normality with the Kolmogorov–Smirnov (K-S) test, and subsequently characterized by their mean and standard deviation (SD). Changes in functional connectivity were assessed at two levels: whole-scalp network summary statistics and pair-wise channel comparisons. Comparisons of global efficiency, psychometric test results, the NASA task load index and the Karolinska sleepiness scale between the exposures were analyzed with a two-tailed paired t-test and reported as mean difference and 95% confidence intervals (CI). Descriptive statistics were generated to characterize the study participants. p-values were regarded significant at p < 0.05, with Bonferroni correction for multiple comparisons.