1.1 Individualized change classification: Problem specification

Suppose a person's RS activity is acquired on two different days. In the period between these two measurements, a person's underlying neurophysiology organization might have changed in a variety of possible ways. However, it is also possible that the person's neurophysiology remained relatively unchanged over this period. These diverse possibilities can be simplified into two categories: (i) change present (i.e., positive instances of neurophysiological change of any kind, denoted as NP+), or (ii) change absent (i.e., negative instances, NP−). This scenario now poses the diagnostic problem of change classification (Figure 1a): How can a person's RS activity from Days 1 and 2 be used to decide whether that person's neurophysiological organization might have undergone a change of any kind (NP+) or not (NP−)?

One simple but erroneous criterion for change classification is the presence/absence of an inter-day difference in RS activity. The diverse variety of possible neurophysiological changes across the brain (i.e., NP+) could produce correspondingly diverse differences in inter-day RS activity. However, the absence of neurophysiological changes (NP−) does not imply the absence of inter-day RS differences due to at least two sources of ‘nuisance’ variability.

Firstly, the repeated measurements on different days can affect inter-day differences in RS activity due to measurement-related factors, as demonstrated by numerous studies of test–retest reliability (Bijsterbosch et al., 2017; Brandmaier et al., 2018; Cox et al., 2018; Noble et al., 2019; Postema et al., 2019). Such spurious activity changes can lead to misclassifications of NP− as NP+ and also NP+ as NP−. Secondly, ‘rest’ is an under-constrained cognitive state that can introduce incidental, idiosyncratic variations in a person's neural state (e.g., Day 1: mind-wandering; Day 2: sleepiness; Day 3: recalling emotional memories) (Benjamin et al., 2010; Diaz et al., 2013; Duncan & Northoff, 2013; Gonzalez-Castillo et al., 2021; Kawagoe et al., 2018). Such cryptic differences in the cognitive state between days could produce differences in RS neural activity even when underlying neurophysiology is unchanged. This can lead to false positives (Type-I errors) where the absence of change (NP−) is misclassified as a neuroplastic change (NP+). Inter-day variability in cognitive states is a particular concern because it is difficult to eliminate and involves true neural activity differences that could confound classification even if measurement-related variability were to be low and highly controlled.

A desirable solution to change classification would be a robust decision rule that enables accurate classification despite the presence of measurement and cognitive state variability between days. ‘Robustness’ is used here in the sense of maintaining performance despite deviations and uncertainties about model assumptions (e.g., Box & Andersen, 1955; Huber, 1981; Xu & Mannor, 2012) and is not used as a synonym for measurement reliability (e.g., Brandmaier et al., 2018). In the current study, we used multivariate pattern analysis (MVPA; Haynes, 2015; Varoquaux et al., 2017) to investigate whether and how such robust decision rules (classifiers) might be obtained. A decision rule's ability to generalize (Bishop, 2006), that is, to classify information from a different setting than the information used to train the decision rule, provided a quantitative measure of robustness.

1.2 Proposed approach: Change classification formulated as cross-day person identification

Identifying a robust decision rule with supervised machine learning involves a critical constraint. For multi-day RS-tracking, the equivalent of one ‘trial’ is the change related to a single pair of RS measurements from different days. Due to practical constraints of RS tracking, only a single trial (i.e., one pair of RS measurements) might be available from a person and the status of this one trial (NP+/NP−) might be unknown and to be determined by classification. Therefore, training a classifier to categorize inter-day differences as either NP+ or NP− using a set of prior examples of these categories from that person would be infeasible. To accommodate this severe constraint, we propose a strategy to use the distinctiveness of individual RS activity as an alternative source of information for change classification.

Our proposed strategy is based on the observation that RS change classification is qualitatively isomorphic to the well-studied problem of RS-based person identification. Numerous studies demonstrate that individual RS activity is highly distinctive to the extent that a person can be identified relative to others solely from their RS activity (Campisi & Rocca, 2014; Finn et al., 2015; see, e.g., Huang et al., 2012; Pani et al., 2020; Valizadeh et al., 2019). In that framework, identification is a form of population inference with a focus on multivariate relationships in a person's RS activity that generalize to samples of the person's own activity but not to the activity of others. These distinctive characteristics are seemingly like the person's biometric signature (or ‘fingerprint’). Crucially, to identify a person across different days, these RS ‘fingerprints’ have to be robust to measurement-related and cognitive state variability (see below). Therefore, we reasoned that procedures used for person identification could be relevant to robust change classification.

The qualitative similarities between change classification and person identification are illustrated in Table 1 (also see Figure 1b). Each row of the table shows the inter-sample relationship of interest, the associated sources of inter-sample differences (measurement, cognitive state, neurophysiological), and the idealized categorization (right column). For idealized change classification (Rows 1 and 2), only the presence/absence of inter-day neurophysiological differences are of relevance to the desired categorization (NP−/NP+), although the other differences are ‘nuisance’ factors to be ignored. This is also the case for idealized cross-day person identification (Rows 3 and 4), where measurement and cognitive state differences are irrelevant for accurate identification. Furthermore, idealized classification of NP− (Row 1) and person S (Row 3) has a similar structure.

For (non-robust) change classification based on simple activity differences, cognitive state differences could lead to a misclassification of NP− (Row 1) as NP+. This nuisance factor can also lead to a misidentification of person S (Row 3) as being not-S. A misidentification could also theoretically occur if the same person undergoes a neurophysiological change across days (Row 3, absent*). This error would be equivalent to the classification of NP+ (Row 2).

Based on this qualitative isomorphism (i.e., between Rows 1-2 and 3-4), a question is whether a same-day classifier trained to identify a particular person (Rows 5 and 6) might, under suitable conditions, be usable for cross-day identification and hence possibly change classification. In prior studies of person identification, same-day classification (Rows 5 and 6), involved samples of a person's RS activity that were obtained from the same RS session of the same day. Thus, the inter-sample differences from the same person (Row 5) can be assumed to have small measurement and cognitive differences, and no neurophysiological differences. However, inter-sample differences from different persons (Row 6) would be associated with both measurement and cognitive state differences, as well as differences in the neurophysiological phenotype. Consequently, a decision rule trained for same-day person identification might not be robust for cross-day person identification unless it represents information about that person's unique neurophysiology in some manner.

We investigated the conditions required for robust inter-day person identification using longitudinal resting EEG activity acquired on 5 consecutive days from healthy, young participants who were assumed to be neurophysiologically stable over this period. A decision rule to identify a person S was trained (using machine learning) to distinguish between (i) examples of RS activity from S on a single day and (ii) examples of single day RS activity from a diverse pool of other individuals (i.e., not-S) (Figure 1b). Each example was a brief 2 s activity sample representing a dynamically variable neural state during an RS session (Calhoun et al., 2014; Hutchison et al., 2013). The decision rule for same day S/not-S classification is used to classify samples of S/not-S from a different day. A robust decision rule should be insensitive to inter-day cognitive variability when a person's neurophysiology is unchanged. This predicted robustness was tested on pseudo-rest states that were designed to challenge cross-day classification with a diverse range of cognitive variation (Figure 2a).

2.1 Participants

Twenty seven healthy volunteers (11 female, age [mean ± SD]: 27.9 years ± 3.4, range: 22–34 years) participated in the study after providing their written informed consent. Participants had normal or corrected-to-normal vision, had no history of neurological or psychiatric disease, were not under medication and had no cranial metallic implants (including cochlear implants). Handedness was not a selection criterion (right handed: 22; left handed: 2; intermediate: 3, based on the Edinburgh Handedness Inventory; Oldfield, 1971). The participants received monetary compensation on completion of all sessions. The study complied with the Declaration of Helsinki and was approved by the Ethics Commission of the Faculty of Medicine, University of Cologne (Zeichen: 14-006).

Datasets from 24 (of the 27) participants were used for statistical analyses (see Section 2.6).

2.2 Apparatus and EEG data acquisition

Stimuli were displayed using the software Presentation (v. 20.2 Build 07.25.18, Neurobehehavioral Systems, Inc.) on an LCD screen (Hanns-G HS233H3B, 23-inch, resolution: 1920 × 1080 pixels). Behavioural responses were recorded with the fMRI Button Pad (1-Hand) System (LXPAD-1 × 5-10 M, NAtA Technologies, Canada).

Scalp-EEG was acquired with a 64-channel active Ag/AgCl electrode system (actiCap, Brain Products, Germany) with a standard 10–20 spherical array layout (ground electrode at AFz, reference electrode on the left mastoid). Electrooculographic (EOG) activity evoked by horizontal and vertical eye movements was recorded with three electrodes (FT9, FT10 and TP10) placed at the left and right lateral canthi and below the left eye respectively. During acquisition, measured voltages (0.1 μV resolution) were amplified, filtered (low cut-off: DC, high cut-off: 250 Hz) and digitized at a sampling rate of 2.5 kHz (BrainAmp DC, BrainProducts GmbH, Germany).

The positioning of the EEG cap was registered with a stereotactic neuronavigation system (Brainsight v. 2.3, Rogue Research Inc, Canada) (details below).

2.3 Experiment protocol and paradigm

Each participant completed five sessions (~40 min each) that were scheduled at the same time on consecutive days (Monday to Friday) at one of three possible time slots: morning (9 am × 6), noon (12 pm × 9) or afternoon (3 pm × 12). For one participant, the fifth session was re-acquired after a gap of 3 days due to technical problems during the scheduled recording. Each session consisted of two RS recordings (RS1 and RS2) interleaved with two non-rest tasks (referred to as Tap and Sequence) in the same fixed order, namely, RS1, Tap, RS2 and Sequence. A schematic of the protocol and the different tasks is shown in Figure 2a.

2.4 Procedure

On the first day, participants received detailed instructions about the different task periods of the experiment. They were familiarized with the number-to-finger mappings required for the Sequence task and practiced the task on a sequence that was different from the one used in the main experiment. Each session included reminders to minimize blinking, maintain fixation at all times during the EEG recording in all task periods, and avoid unnecessary movements of the fingers, head and body. At the start of each session, participants completed the Positive and Negative Affect Schedule (PANAS) (Watson et al., 1988) and brief questionnaires about their caffeine consumption on that day and the amount and quality of sleep on the previous night.

We used a prospective strategy to minimize inter-day variation in the positioning of the EEG cap with a spatial registration procedure on each day. Using a stereotactic neuronavigation system, the participant's head was registered to the Montreal Neurological Institute (MNI) space using standard cranial landmarks. The positions of five selected electrodes along the midline and lateral axis (AFz, Cz, POz, C5, C6) were then registered in this space. The electrode locations from the first day were used as a spatial reference for the remaining sessions. On each subsequent session, the cap's position was adjusted to align these selected electrodes to their reference locations. Due to scheduling constraints, this spatial registration procedure was not performed for seven participants.

The application of electrode gel followed after cap positioning. Skin-electrode impedance was brought below 10 kΩ before starting the recording.

Recordings were acquired in a light-dimmed and acoustically shielded EEG chamber. Participants were seated at a comfortable chair with their heads stabilized with a chinrest in front of the computer screen at a viewing distance of ~65 cm. The response pad was placed in a recess under the table to prevent visibility of the hands during the task periods. Participants were monitored during the recording via a video camera to ensure that they maintained fixation, minimized eye-blinks and stayed awake.

2.5 EEG preprocessing

The EEG data were preprocessed using the EEGLAB software (Delorme & Makeig, 2004) and custom scripts in a MATLAB environment (R2016b, MathWorks, Inc., Natick, MA).

2.6 Data quality assessment

Preprocessing resulted in 135 session datasets (27 participants × 5 days). For inclusion in the final analysis, each participant had to have completed the first three of the four tasks on all sessions and have at least four (out of 5) session datasets that met the following data-quality criteria. We required a preprocessed session dataset to have (i) less than seven rejected channels, (ii) ≥ 90 artefact-free epochs from both RS periods (i.e., RS1 and RS2) and (iii) ≥ 90 artefact-free epochs from the available resting-matched conditions (i.e., TapWait and SeqWait). Note that the number of epochs for TapMov and SeqMov were necessarily ≤ 60 as each task only had 60 response periods of 2-s duration. To maintain uniformity across participants, final analyses were performed only on the best four of the five session datasets from each participant. If all five session datasets were of high quality, the first day's dataset was excluded as it might involve effects of initial familiarization.

Datasets from 24 out of 27 participants met the above data-quality criteria: 18 (of 24) had completed all four task periods on each session while the remaining six (of the 24) participants had completed only the first three (of the four parts). Only two of the 24 participants required the use of the first day's dataset. To maximize the use of the available data, analyses involving only RS1 and RS2 included data from 24 participants, while analyses of the non-rest tasks used data from 18 participants.

For the 24 participants, the mean number of epochs per day per participant for RS1 was 137.81 (min: 125.7, max: 146.5, SD = 5.28, min/day = 96), and for RS2 was 138.11 (min: 116.5, max: 148.0, SD = 6.89, min/day = 91). For the 18 participants used to analyse the non-rest tasks, the mean number of epochs per day per participant for TapWait was 160.6 (min: 146.0, max: 175.0, SD = 7.23, min/day = 129), and for SeqWait was 172.29 (min: 156.5, max: 186.5, SD = 9.13, min/day = 129). The corresponding values for lower for TapMov = 55.62 (min: 49.0, max: 59.0, SD = 2.62, min/day = 41), and for SeqMov = 56.74 (min: 53.25, max: 59.5, SD = 1.78, min/day = 43).

2.7 Classifier specification

Each epoch was a 2-s sample of the ongoing activity from one person (of 24) on one specific day (of 4), although in a particular task state (of six possible states: true rest [RS1, RS2], pseudo-rest [TapWait, SeqWait] and non-rest [TapMov, SeqMov]). For our analyses, the basic classification problem was whether an epoch's activity could be used to identify (i.e., classify) that epoch's origin either by (i) a person's identity (using a multi-class classifier) or (ii) task state within the same person (using a standalone binary classifier).

2.8 Multi-class classification

Numerous prior studies demonstrate that RS activity can serve as a ‘fingerprint’ for person identification (Campisi & Rocca, 2014; Finn et al., 2015; Huang et al., 2012; Pani et al., 2020; Valizadeh et al., 2019). Although our focus was not on the neural basis of individual differences and trait identification (Demuru et al., 2017; Finn et al., 2017; Gratton et al., 2018; Smit et al., 2005, 2006), a person identification approach, using multi-class classifiers, provided a convenient technical platform for our test of individual-specific change classification.

2.9 Evaluation of classifier robustness

The robustness (or generalization) of person identification with a particular classifier was evaluated by the quality of identification decisions on test samples with a different origin from the samples used to obtain the classifier (i.e., the training samples). This evaluation was organized into two schemes based on whether the training and test samples belonged to the (i) same day versus a different day (Figure 3a) and the (ii) same task state versus a different task state.

2.10 Weights and normalized weights

From this expression, a feature weight that differs from zero irrespective of sign (i.e., |w_i| > 0) is an indicator that the corresponding feature was relevant for classification even if only indirectly (Haufe et al., 2014; Schrouff & Mourao-Miranda, 2018). However, the relative contribution of a feature i to the final decision value also depends on |w_iP_i|. For features i and j, the weight |w_i| might be greater than |w_j|, while the sample power |P_i| might be less than |P_j.|. Consequently, neither the raw absolute weights nor power is an unambiguous guide to the relative influence of features i and j on the classification decision. Therefore, we defined a feature i's unit weight ${\overline{w}}_i$ as the idealized weight value such that ${\overline{w}}_i{P}_i=1$. The normalized weight was thus defined as the ratio $w_i/{\overline{w}}_i$, which was effectively equal to w_iP_i.

A feature's characteristic weight was obtained by averaging the feature's weight over all decision rules consistent with a particular scheme (e.g., ^AI_p → ^AI_q) (Figure 3a). Similarly, a feature's characteristic power was the grand average of the power across samples and days. The absolute normalized weights (i.e., |wP|) were z-scored within each band for each subject to retain information about band-specific, inter-feature weighting differences.

2.11 Statistical analysis

Statistical tests were performed using the pingouin package (version 0.3.2) (Vallat, 2018) and MATLAB Statistics toolbox. The random chance accuracy for the multi-class and standalone binary classifier was 50%, and accuracy deviations from random chance were evaluated with one-sample t tests. Correlations between individual accuracy values were evaluated using Spearman's rank correlation due to the focus on relative ordering rather than a strict cardinal relationship. Tests with a p value < alpha = .05 were deemed to be statistically significant. For tests involving multiple comparisons, p values were evaluated against a Bonferroni-corrected alpha threshold. Due to the sequential relationship between the different multiclass classification schemes, the planned tests on the same-day accuracy (CV) and cross-day accuracy were evaluated at an alpha threshold of .05. However, tests on 2- and 3-day accuracy were evaluated at a threshold of alpha = (.05/2).

For plots depicting mean values at different levels of a single factor, error bars indicate the standard deviation (SD). For plots depicting the effects of multiple factors, error bars displaying the within-subject standard error (s.e.m.) (O'Brien & Cousineau, 2014). The type of error bar used is explicitly noted in the figure caption.

3.1 Face validity of individual power spectra

Our investigation assumed that an individual's power spectrum at rest can systematically (i) differ between days and also (ii) differ from the spectra of other individuals. We first confirmed the face validity of these assumptions in our data.

The presence of structured inter-individual differences during RS1 was qualitatively evident in the mean (full) power spectrum at different channels (Figure 4a) before its reduction to the minimal description used for the classification analyses. As shown for one example individual S₁, individual power spectra had a similar form across channels with a higher power in the δ and α bands and a higher overall power in the posterior and anterior channels relative to the central channels. The diversity of pairwise differences between individual spectra highlights the difficulty of representing an individual's unique properties. For example, the combination of channels and frequencies (i.e., features) at which S₂ and S₃ showed prominent differences were not the same features at which S₂ differed from S₅. Nonetheless, the required decision rule to identify S₂ was a single feature configuration capable of distinguishing S₂ from all others while allowing S₂ to be robustly re-identified across days.

Systematic inter-day differences were evident from the dissimilarity between samples from all participants and all days (90 samples per participant per day) (Figure 4b). The dissimilarity between any two samples was described by their correlation distance (= 1 − r, where r is the Pearson's correlation coefficient) (Diedrichsen & Kriegeskorte, 2017; Dimsdale-Zucker & Ranganath, 2019; Pani et al., 2020). For all 24 participants, the mean dissimilarity between samples from the same day was lower than between samples from different days (cross-day) (t₂₃ = −6.74, p < .0001). However, the dissimilarity between same-day and cross-day samples varied from person to person suggesting their possible confusability with samples from other individuals. This was the critical issue to be resolved with an appropriate decision rule, to be identified using machine learning.

3.2 Individual identification from RS activity within and across days

3.2.1 High same-day accuracy but reduced cross-day accuracy of individual decision rules

To identify a person from a 2-s sample of RS activity with an ensemble classifier, a decision rule was numerically estimated to represent each person's unique RS characteristics. The decision rules estimated for each day could identify each person (of 24) from a sample acquired on the same day (i.e., according to the scheme ^RS1I_p → ^RS1I_p) with a mean cross-validated (CV) accuracy of 99.98 ± 0.04% (mean ± SD) that was significantly larger than the theoretically expected accuracy for random guessing (> 50%: t₂₃ = 5596.13, p < .00001) (Figure 5a, Table A.1). However, for longitudinal tracking, a key demand is that decision rules from 1 day should identify a person from samples acquired on a different day (i.e., ^RS1I_p → ^RS1I_q). The same-day decision rules identified individuals across days with a mean accuracy of 92.10% ± 6.8% that was higher than random chance (t₂₃ = 30.14, p < .00001) but less accurate than same-day identification by ~8% (paired t₂₃ = 5.64, p = .00001).

The confusion matrix (Figure 5b) of how individuals were misclassified during cross-day (1-day) identification revealed four clusters of individuals who were confused with each other. Notably, the individuals with the lowest cross-day accuracies (namely, S₂, S₁₁, S₁₅ and S₂₄) belonged to different clusters rather than being solely confused with each other. The clustering of misclassified individuals suggested that errors in identifying an individual S_X were due to a combination of (i) changes to S_X's RS activity between days (i.e., false negatives) and (ii) changes to other individuals who were then misclassified as S_X (i.e., false positives). Nevertheless, the increased errors in individual identification illustrate the challenge of NP+/NP− decisions. Errors in identifying a person S_X across days seemingly imply that S_X's unique identifying characteristics had changed across days even though the individuals here were unlikely to have changed in their underlying neurophysiology over the 5-day testing period.

3.3 Information organization in RS activity for individual identification

The hypothesized configuration in RS activity was suggestive of a multivariate relationship between distributed features. However, the accuracy relationships described above do not indicate whether such a distributed configuration was necessary to enable individual identification. Therefore, we evaluated the information organization required for individual identification.

3.3.1 Low cross-day identification with information from only one frequency or one location

Each sample was a snapshot of RS activity described by 305 informational features (5 bands × 61 channels). To test the informational role of these different features, we evaluated whether identification comparable to the full feature set was possible with subsets of features that were defined either by frequency band (i.e., mono-band sets) or spatial location (i.e., mono-location sets).

Each mono-band feature set (B_f) consisted of features from one frequency band f at all 61 channels. For all five mono-band sets (Figure 6a, Table A.2), same-day identification had a mean accuracy greater than 95%. However, the size of the cross-day loss in accuracy was band-dependent and ranged from ~14% for B_α to nearly ~32% for B_δ (ANOVA, type [CV, 1-day] × band [B_δ, B_θ, B_α, B_β1, B_β2], type * band: F_4,92 = 24.83, p < .00001; type: F_1,23 = 232.11, p < .00001; band: F_{4, 92} = 40.30, p < .00001). The divergence in cross-day losses for B_α and B_δ was striking as these two bands have a characteristically higher power relative to the other bands (Figure 4). Training with multi-day aggregation (Figure 6b) increased cross-day accuracy by differing amounts for each band by, for example, +10% for B_β2 but only +6% for B_δ (ANOVA, band [B_δ, B_θ, B_α, B_β1, B_β2} × type [1-day, 2-day, 3-day], type * band: F_{8, 184} = 9.19, p < .00001; type: F_{2, 46} = 146.02, p < .00001’; band: F_{4, 92} = 43.13, p < .00001). However, even with 3-day aggregation, the residual difference between cross-day and same-day accuracy (minimum: ~7% for B_α, maximum: ~26% for B_δ) was larger than the ~2% difference with the full feature set.

Each mono-location feature set (L_z) consisted of 50 features (5 bands × 10 channels) in the spatial zone z (Figure 2a). The mean same-day accuracy was greater than 95% for all mono-location feature sets (Figure 6c, Table A.2). However, the mean cross-day (1-day) accuracy showed reductions of ~12%–16% for all locations (ANOVA, type [CV, 1-day] × location [L_F, L_FC, L_CP, L_PO], type * location: F_3,69 = 3.77, p = .015; type: F_1,23 = 108.91, p < .00001^,; location: F_{3, 69} = 5.45, p = .0020]. The mean cross-day accuracy for the fronto-central (L_FC) and centro-parietal (L_CP) sets were marginally higher than for the parieto-occipital (L_PO) and frontal (L_F) sets. This zonal accuracy difference was notable as the mean power for all bands was typically higher over the posterior and anterior channels than the centrally located channels (Figure 4a). Aggregation increased cross-day accuracy by ~6% for all four location sets (Figure 6d) (ANOVA, location: [L_F, L_FC, L_CP, L_PO] × type [1-day, 2-day, 3-day], type * location: F_{6, 138} = 2.07, p = .06; type: F_{2, 46} = 115.38, p < .00001; location: F_{3, 69} = 4.79, p = .0043). Nevertheless, the residual ~7%–10% loss in cross-day accuracy was larger than with the full feature-set.

In summary, all the mono-band and mono-location sets contained sufficient information to enable same-day identification with nearly error-free accuracy. However, this information had a low day generality. Even with aggregation, these feature sets had a lower cross-day accuracy than the full feature-set that combined these feature sets together. This divergence suggests that the higher cross-day robustness with the full feature set involves a role for multivariate relationships between different frequency bands (i.e., unlike the mono-band subsets) at spatially distributed channels (i.e., unlike the mono-location subsets). To assess how this multi-feature configuration might be organized, we evaluated the pattern of weights associated with the different features of the full feature set.

3.3.2 Concentration of high-consistency informative features at fronto-central and occipital zones

Each individual's decision rule was defined by the configuration of weights assigned to the different features. Because a decision rule uniquely identifies a person, the feature weights that define a person's decision rule would need to be different for that of all others. Despite these weighting differences, certain features might nevertheless be informative across individuals. To identify these features, we evaluated the inter-individual consistency in the influence of different features on the identification decision. A feature's influence on the individual's decision rule was quantified by the feature's normalized weight (to correct for inter-feature power differences) that was then z-scored (within each frequency band) to retain inter-feature relevance differences.

Figure 7 shows the topographic distribution of these high-consistency features of the full feature-set with mean normalized, z-scored weights (averaged across individuals) that were significantly greater than zero (see Figure S1 for corresponding non-normalized [raw] weights). At corrected thresholds (see t values in Figure 7, lower panels), the features associated with all frequency bands except the δ band contained at least one high-consistency feature. Rather than having an idiosyncratic organization, the high-consistency features were concentrated at distinctive zones in each frequency band.

In B_θ, there was a concentration of high consistency features at CP1 and C3, with the addition of CP3 with aggregation. There was a similar, although weaker, concentration of consistent features at corresponding channels over the right hemisphere. Showing a similar spatial organization, the high-consistency features in B_β1 showed a striking bilaterally symmetric configuration along the transverse midline at channels C3, Cz and C4 with an aggregation-modulated role for CP6 and T7 (and possibly T8). This similarity in organization was notable since the frequency ranges of the θ band (4–7.5 Hz) and β₁ (14–22.5 Hz) were not contiguous and were separated by the α band.

Unlike this central concentration of features in B_β1 and B_θ, the features in B_α contained a single, strongly consistent feature in the occipital zone at PO3. At uncorrected thresholds, there were other distributed features across the scalp that were weakly consistent for both 1- and 3-day identification, namely, at AF3, C3, P8 and O2. Similarly, the features of the high-frequency β₂ band (i.e., B_β2) only had a single consistent feature at P1 with a diffuse distribution of consistent features at uncorrected thresholds.

In general, the distribution of high-consistency features was by itself not a simple indicator of their contribution to cross-day accuracy. For example, the relative number of high-valued (raw) weights in the different bands and spatial locations had a low correspondence to relative accuracy of cross-day identification based solely on the mono-band/location subsets (see Figure S2). Nevertheless, the organized distribution of high-consistency features at channels over the sensorimotor cortex and the occipital cortex was prima facie support for an individual-specific configuration with a basis in neurophysiological constraints. These high-consistency zones were of particular relevance to the relationship of RS1 to the non-rest task states where the power over the sensorimotor and occipital zones was expected to differ from RS1.

3.4 Relationship of rest to non-rest states

The behavioural demands during TapMov and SeqMov were designed to modulate the cognitive states during the TapWait and SeqWait periods and produce neural activity deviations from RS1 in the absence of behavioural differences. Furthermore, the Tap and Sequence tasks were designed to elicit neural states that varied between days for Sequence (low cross-day similarity) but remained constant for Tap (high cross-day similarity). We sought to first explicitly verify that such deviations from RS1 were indeed present. Note that all analyses of Tap and Seq states were performed in a subgroup of N = 18 participants (see Section 2).

3.4.1 Neural activity during Tap and Sequence verifiably deviates from RS1

The inter-day changes in behaviour during the TapMov and SeqMov periods were consistent with the experimental assumptions (Figure 8a). During TapMov, the mean number of button presses during the cued 2-s period (~10–11) remained effectively constant across days (one-way ANOVA, F_{4, 68} = 0.50, p = .73). In contrast, during SeqMov, the mean number of button presses increased from ~7 on the first day to ~11 on the fifth day (one-way ANOVA, F_{4, 68} = 40.75, p < .00001). This inter-day change in motor performance in SeqMov was systematically different from TapMov as confirmed by the statistically significant interaction in an ANOVA with factors (condition [TapMov, SeqMov] × days [D1,…, D5]; condition * days: F_{4, 68} = 21.00, p < .00001; condition: F_{1, 17} = 5.73, p = .03; days: F_{4, 68} = 21.53, p < .00001).

The neural state during the movement period (TapMov, SeqMov) showed typically expected dynamic states (Figure 8b, Supplementary Methods 2). Changes in the mean β power at channel C4 (contralateral to the moved fingers) were in line with the event-related de-synchronization/synchronization (ERD/ERS) phenomenon for repetitive movements (Alegre et al., 2004; Cassim et al., 2000; Erbil & Ungan, 2007; Pfurtscheller & da Silva, 1999), namely, a power reduction at the onset of movement execution (i.e., ERD) with an increase after the termination of all movements (i.e., ERS). Furthermore, the β power changes at Oz showed a task-dependent neural response consistent with differing visual stimulation, that is, an increase for TapMov (blank screen) but a decrease for SeqMov (image depicting the sequence). These movement-vs-wait differences were validated in the samples used for classification. A within-subject binary classification of TapWait versus TapMov had a mean cross-validated accuracy of 85.91 ± 7.23% (>50%: t₁₇ = 21.06, p < .00001); and SeqWait versus SeqMov had a mean CV accuracy of 94.58 ± 3.20% (>50%: t₁₇ = 59.02, p < .00001).

The critical verification for our study was the relationship between RS1 and the pseudo-rest states (TapWait, SeqWait). Samples from TapWait and SeqWait were distinguishable from RS1 on the same day with high cross-validated accuracy (RS1 vs. TapWait: 88.28 ± 5.70%; RS1 vs. SeqWait: 95.12 ± 3.74%) (Figure 8c, left panels, Table A.3). However, the cross-day accuracy (without aggregation) for both RS1 versus TapWait (62.91 ± 6.44%) and RS1 versus SeqWait (67.79 ± 8.53%) was substantially lower than the same-day accuracy by more than ~25%. Nevertheless, the cross-day accuracy for RS1 versus SeqWait was marginally higher than for RS1 versus TapWait with increasing aggregation (ANOVA: condition [RS1 vs. TapWait, RS1 vs. SeqWait] × type [1-day, 2-day, 3-day], condition * type: F_{2, 34} = 6.22, p = .005; condition: F_{1, 17} = 8.37, p = .01009; type: F_{2, 34} = 38.89, p < .00001).

TapMov and SeqMov were also distinguishable from RS1 on the same-day with high (cross-validated) accuracy (RS1 vs. TapMov: 93.56 ± 4.12%; RS1 vs. SeqMov: 97.81 ± 1.76%) (Figure 8c, right panel, Table A.3). Similar to the wait periods, the cross-day accuracy for RS1 versus SeqMov was higher than for RS1 versus TapMov across aggregation levels (ANOVA: condition [RS vs. TapMov, RS1 vs. SeqMov] × type [1-day, 2-day, 3-day], condition * type: F_{2, 34} = 0.61, p = .55; condition: F_{1, 17} = 30.91, p = .00003; type: F_{2, 34} = 69.47, p < .00001).

The above findings verified the neural activity differences in the task states in Tap and Sequence to each other and to RS1. Crucially, the structure of the same-day differences had a low cross-day generality.

3.4.2 Robust identification of individuals from Tap and Sequence activity within and across days

The above differences between task states and RS1 raised the issue of whether the task-related functional states also disrupt the information that enables individual identification with RS1. To assess this possibility, we evaluated whether the different Tap and Sequence task states contained sufficient information for person identification in a same-task classification scheme (i.e., with the scheme ^XI_p → ^XI_q for task X) (Figure 8d).

The same-day accuracy for both TapWait and SeqWait was ~99% (Figure 8d, left panels, Table A.1). The mean cross-day accuracy (without aggregation) for TapWait (92.58 ± 6.39%) was lower than its corresponding same-day accuracy by only ~7% (t₁₇ = 4.92, p = .00013). Similarly, for SeqWait, the mean cross-day (1-day) (93.67 ± 7.35%) accuracy was lower than the same-day accuracy by ~6% (t₁₇ = 3.65, p = .00197). Furthermore, the effect of aggregation on mean cross-day accuracy for TapWait and for SeqWait was statistically indistinguishable (ANOVA: condition [TapWait, SeqWait] × type [1-day, 2-day, 3-day]; condition * type: F_{2, 34} = 0.88, p = .42; condition: F_1,17 = 1.35, p = .26; type: F_{2, 34} = 21.30, p < .00001).

Despite the deviations of TapMov and SeqMov along both the behavioural and cognitive dimensions of rest and their differences with each other, the accuracies of individual identification across days for TapMov and SeqMov were greater than 90% for all levels of aggregation and were not statistically distinguishable from each other (Table A.1, Figure 8d, right panels) (ANOVA: condition [TapMov, SeqMov] × type [1-day, 2-day, 3-day]; condition * rype: F_{2, 34} = 0.86, p = .43; condition: F_{1, 17} = 1.26, p = .28; type: F_{2, 34} = 14.50, p = .00003).

Thus, individual identification was robustly possible in the task states despite their differences to RS1. Furthermore, the identification accuracy was similar between the Tap and Seq states despite their functional differences. Two further lines of evidence supported the possibility that these similarities were based on common task-independent properties. The spatial distribution of high-consistency features for these states (Figure 9a, Figure S3) exhibited a striking qualitative similarity to each other as well as to the corresponding distribution for RS1 (Figure 7). Additionally, the individual cross-day (1-day) accuracy in these task states showed a striking correlation to the corresponding cross-day accuracy in RS1 (Figure 9b) (threshold: p < .05/4; TapWait: r[17] = .882, p < .00001; SeqWait: r[17] = .635, p = .00466; TapMov: r[17] = 0.75, p = .00034; SeqMov: r[17] = .653, p = .00329). Thus, the inter-individual relationships revealed by the errors in cross-day classification during RS1 (Figure 5b) seemingly extended to these non-rest states as well. We next turned to a formal assessment of this cross-task relationship.

3.5 Robust generalization of rest-based decisions to cross-task individual identification

If person identification with RS1 was based on a neural configuration related to an individual's neurophysiological state, then identification should be possible despite cognitive state variations. Therefore, decision rules trained on RS1 should be capable of accurate person identification with samples acquired from the pseudo-rest states (TapWait and SeqWait) and the movement states (TapMov and SeqMov).

We used the cross-task scheme ^RS1I_p → ^XI_q to test the invariance of RS1-based identification to inter-day cognitive state variations (i.e., task states X) (Figure 10a, Table A.4). Increasing deviations from RS1 solely due to cognitive state differences (X = [RS1, TapWait, SeqWait]) did not produce comparable, statistically distinguishable reductions in mean identification accuracy (RS1: 92.79 ± 6.76%, TapWait: 91.90 ± 6.46%; SeqWait: 90.81 ± 7.09%) (one-way ANOVA, F_{2, 34} = 2.06, p = .14). However, increasing deviations from RS1 due to cognitive and behavioural state differences (X = [RS1, TapMov, SeqMov]) produced significant reductions in identification accuracy most notably for SeqMov (TapMov: 88.79 ± 7.57%; SeqMov: 83.85 ± 10.35%) (one-way ANOVA, F_{2, 34} = 14.07, p = .00004).

To disentangle the role of cross-task from cross-day effects, we compared cross-task (^RS1I_p → ^XI_q) and same-task identification (^XI_p → ^XI_q) across days (Tables A.4 and  A.1 respectively). For the pseudo-rest states (X = [TapWait, SeqWait]), cross-task accuracy with RS1 decision rules produced a small but statistically significant reduction relative to same-task identification (ANOVA, train [RS1, Same] × condition [TapWait, SeqWait], train * condition: F_1,17 = 4.14, p = .06; train: F_1,17 = 10.02, p = .00566; condition: F_{1, 17} = .00001, p = 1.00). The cross-task accuracy reduction was significantly larger for the movement states (X = [TapMov, SeqMov]) with a larger loss for SeqMov (ANOVA, Train [RS, Same] × condition [TapMov, SeqMov], train * condition: F_1,17 = 9.15, p = .00764; train: F_1,17 = 43.94, p < .00001; condition: F_{1, 17} = 2.51, p = .13).

To disentangle the role of day specificity in ^RS1I_p → ^XI_q, we used multi-day aggregation (^RS1I_p ∘ ^RS1I_q … → ^XI_r). Although aggregation reduced day specificity with RS1 (Figure 5), this could have been achieved by increasing specificity to the properties of RS1. Such an ‘overfitting’ to RS1 (i.e., increased task specificity) might lower the accuracy of cross-task identification. Alternatively, aggregation could have reduced both day and task specificities and thus increase the accuracy of cross-task identification. Consistent with this latter possibility, aggregation increased cross-task accuracy to the pseudo-rest states (TapWait, SeqWait) in a comparable manner to same-task accuracy (Figure 10b) (ANOVA: condition [RS1, TapWait, SeqWait] × type [1-day, 2-day, 3-day]; condition * type: F_{4, 68} = 0.52, p = .72; condition: F_{2, 34} = 2.44, p = .10; type: F_{2, 34} = 21.63, p < .00001). This was particularly striking because aggregation (i.e., related to day specificity) produced a relatively larger increase in cross-task accuracy than a change in task specificity. Following aggregation, the mean residual cross-task/cross-day accuracy loss relative to same-task/cross-day identification with RS1 was only ~3%. Aggregation also increased cross-task accuracy to the movement states (TapMov, SeqMov) (ANOVA: condition [RS1, TapMov, SeqMov] × type [1-day, 2-day, 3-day]; condition * type: F_{4, 68} = 1.35, p = .26; condition: F_{2, 34} = 13.04, p = .00006; type: F_{2, 34} = 29.33, p < .00001). Following aggregation, the mean residual cross-task/cross-day difference was less than ~10% for the movement states.

Similar to the same-task correlations described above (Figure 9b), the individual cross-task (1-day) accuracy in each of these task states showed a statistical significant correlation to the corresponding cross-day accuracy in RS1 (Figure 10c) (threshold: p < .05/4; TapWait: r[17] = .948, p < .00001; SeqWait: r[17] = .771, p = .00018; TapMov: r[17] = .897, p < .00001; SeqMov: r[17] = .631, p = .00503). The correlation coefficients were particularly high for both Tap states (TapWait and TapMov) as compared with the Seq states (SeqWait and SeqMov). Furthermore, the scatter plots suggested that the relatively lower cross-task/cross-day accuracy for SeqMov was driven by the low generalization of a few individuals.

In summary, decision rules trained on RS1 on a single day could identify individuals from samples from states that verifiably differed from RS1 to differing extents. Importantly, aggregated training solely on RS1 lead to increases in identification accuracy on samples from these non-rest task states. Unlike the full-feature set, applying the cross-task scheme ^RS1I_p → ^XI_q to the mono-band and mono-location feature sets produced large accuracy reductions (Figure S4). This confirmed that the conjoint role of features from more than one frequency band and spatial zone was crucial to obtain high cross-task/cross-day identification accuracy. Taken together, the cross-task/cross-day robustness of person identification with a distributed feature-set was consistent with the properties of a configuration constrained by individual neurophysiology, that is, a critical demand for change classification.

4.1 Change evaluation versus change-informed representation

A plausible alternative approach to change classification is to directly evaluate inter-day RS activity, for instance, with a dissimilarity measure (as in Figure 4b) or an assessment of inter-day variability within a test–retest reliability framework (Bijsterbosch et al., 2017; Cox et al., 2018; Noble et al., 2019; Postema et al., 2019). The obtained score could then be converted into a change classification decision, for example, a suitably high inter-day reliability (or similarity) might suggest an NP− categorization (i.e., no neurophysiological change), whereas a low reliability (or similarity) would suggest an NP+ categorization. However, such evaluation-based approaches would be incomplete without specifying how individual RS activity is to be represented for this evaluation, as illustrated by the following analogy to object recognition.

Consider images depicting the same object X from Days 1 and 2 (Figure 11a) represented by filled pixel locations (i.e., features). A simple measure of inter-day reliability is whether the filled pixels on Day 1 are also reliably filled on Day 2. Scenario A would have a high feature-level reliability as most filled pixels on Day 1 are also filled on Day 2 (suggesting NP−), whereas Scenario B would have a low reliability (suggesting NP+). However, these change classification decisions are faulty inferences about the overall shape of the depicted object. In Scenario A, the object's shape differs between days (i.e., analogous to NP+), although in Scenario B, the object's shape is unchanged despite a large change in orientation (i.e., analogous to NP− as with TapWait, SeqWait, TapMov, SeqMov). The faulty inference can be attributed to the representation of the object (i.e., list of filled pixel locations) as its format is uninformed by the possible types of change. However, an ideal change-informed representation of object X's shape would be invariant to irrelevant rotations as in Scenario B (analogous to cognitive state variation), while being sensitive to true shape changes as in Scenario A (analogous to neurophysiological change) (Figure 11b).

Similarly, our approach was based on the view that RS change classification requires a change-informed representation of individual's RS activity for evaluation. Ideally, this representation would be sensitive to neurophysiological changes (NP+) but invariant to incidental cognitive variation (NP−). In analytical terms, we assume that there is a one-to-many mapping between an individual's neurophysiological phenotype and the multiple functional neural activity states (including at rest) that are uniquely configured by that phenotype (i.e., NP → [activity states]). Hypothetically, this one-to-many mapping (i.e., NP → [activity states]) could be analysed to extract a one-to-one mapping between an individual's neurophysiology to a unique configuration of constraints shared by the many activity states (i.e., NP → [constrained configuration] → [activity states]) (Figure 11b). Such a constraint configuration could serve as a change-informed representation as it would be invariant to incidental cognitive variation (NP−) but sensitive to neurophysiological change (NP+).

Such a change-informed representation might not be known a priori. In such a scenario, training a decision rule for person identification (Figures 1c and 3a) could serve as a data-driven procedure to select such a change-informed representation. The term ‘representation’ is used here to refer to how information is carried by a configuration of features (e.g., by the specific assignment of weights to different features) rather than in the sense of ‘feature selection’ (Guyon et al., 2002; Guyon & Elisseeff, 2003), that is, to find a subset of features that carry relevant information.

4.2 Person identification to select change-informed representations

Samples of same-day activity from (1) other individuals (not-S) and (2) person S serve as examples of the possible types of change, namely, NP+ and NP−, respectively. Therefore, training a decision rule on these examples (using machine learning) was a mechanism to select a putative change-informed representation.

As a simple analogy, in Figure 11a, object X's shape can be distinguished from that of object Y on Day 1 based on a few critical pixels (circled). In Scenario A, this pixel subset from Day 1 is unfilled on Day 2 implying that X's identity on Day 2 is now potentially confusable with object Y, which in turn suggests a possible cross-day change in shape (NP+). Thus, inter-individual comparisons can help to select critical feature relationships to represent an object X's identity, which is based on the object's shape. The critical pixels are unchanged across days in Scenario B suggesting an unchanged identity and hence shape (NP−). However, this representation suffers the limitation of being uninformed about rotations, as described above with the full list of filled pixels. Therefore, selecting a representation that was robust to incidental change was crucial.

Despite using an analogy of RS activity to a static object, training for person identification on same-day activity involved an assumption about dynamics and timescales. Each same-day measurement was segmented into 2-s non-overlapping activity samples. This inter-sample variability on short timescales (i.e., between the samples acquired within seconds/minutes of each other on the same day) was assumed to contain information about how RS activity could change in the absence of neurophysiological change (i.e., NP−). Therefore, successful cross-day identification was predicated on whether inter-sample differences on short timescales on the order of seconds were informative about inter-sample differences on long timescales (i.e., hours and days apart). Using moment-to-moment variability to ‘account’ for incidental RS differences was critical to bypass limits on available information. The cognitive state during rest measurements is related to experimental context and instructions (Duncan & Northoff, 2013; Kawagoe et al., 2018). However, beyond the assumption that participants were awake, we did not model the participant's cognitive state, for example, using participant's self-reported assessments of their cognitive state during the RS measurement, sleepiness or coffee consumptions (Diaz et al., 2013; Guerra-Carrillo et al., 2014).

The individuality of RS activity has been studied with a variety of objectives, such as biometric identification (Campisi & Rocca, 2014; Gui et al., 2014; Valizadeh et al., 2019) and general questions related to the neural basis of individual differences and trait-identification (Demuru et al., 2017; Finn et al., 2017; Gratton et al., 2018; Smit et al., 2005; Smit et al., 2006). Consistent with these studies, our results also demonstrate the high distinctiveness of individual RS activity. Person identification was possible significantly above random chance from two-second snapshots of the power spectra at rest within the same day, as well as across days and tasks (Tables A.1, A.2 and A.4). However, in the current study, person identification served as a procedure to select change-informed representations. Hence, it was not sufficient to identify a person, for example, with a (dis)similarity-based measure that does not provide such a representation (e.g., Finn et al., 2015). Furthermore, the modulation of cross-day accuracy (i.e., accuracy loss) was a key indicator and classification above random chance was not itself informative about cross-day change. For these reasons, a machine learning approach was valuable as a principled method to obtain an explicit representation of the individual to be used for change classification.

4.3 Trade-offs of data-driven selection of representations

Selecting representations based on a functional criterion (i.e., the ability to distinguish S from not-S) involved certain trade-offs.

There was no guarantee that the selected representations would be linked to individual neurophysiology rather than arbitrary day-specific properties. Even in the toy example (Figure 11a), the objects differ in colour (red = object X, blue = object Y), but using this feature to assess inter-day change would be uninformative about shape changes (as in Scenario A). Therefore, establishing the empirical soundness of this approach was critical. Because cognitive variability across days is difficult to establish, the high identification accuracy with RS1 could have been attributed to highly motivated and instruction-compliant participants rather than the neural characteristics of the rest state. However, the Tap and Sequence tasks provided verifiable within-subject examples of states that deviated from rest in order to assess the generality of RS-based inferences. Additional validity checks were provided by the battery of empirical tests, for example, the effects of aggregated training; assessing day versus measurement-specificity; and the weight distribution. An important boundary condition is that high same-day accuracy for person identification is not sufficient to obtain a change-informed representation as shown by the poor cross-day accuracy with mono-band/mono-location feature subsets (Figures 3 and 6, Figure S4).

As an individual's identity is defined relative to other individuals in the studied group, an individual's representation would vary depending on the diversity of the group and properties of the most-similar individuals (as illustrated by the confusion matrix and inter-individual clustering in Figure 5). Furthermore, features shared by all individuals would be excluded. For example, in a study of the heritability of individual RS-connectivity properties with magnetoencephalography (MEG) (Demuru et al., 2017), the explicit removal of connectivity characteristics shared by all individuals in the group significantly improved individual identification. Thus, downweighting the role of shared features (explicitly or implicitly) would prevent changes to these shared features from being detected.

Furthermore, selecting representations based on a functional criterion enables considerable generality as the criterion does not specify the kind of information being represented. For instance, in our study, each activity sample was defined by the power in the canonical frequency bands. However, each sample could alternatively be defined by, for example, the dynamic connectivity estimated from the oscillatory phase (Bonkhoff et al., 2021; Calhoun et al., 2014; Rosjat et al., 2018; Rosjat et al., 2020). Exploring such extensions to other forms of information and information obtained from other imaging modalities is a key topic for future studies.

4.4 Outlook

Our results support the technical feasibility and potential value of RS change classification to support the use of RS to track neuroplastic change. Notwithstanding its trade-offs, person identification suggests a powerful and convenient strategy to select appropriate change-informed representations to support change classification. We assumed that individuals in the studied group did not undergo extensive plastic changes. If individual identification was not possible with longitudinal RS even with such a group of healthy individuals over a period of 5 days, then the merits of using RS as a tracking indicator would seem to require critical re-evaluation especially for tracking over longer periods of time and with populations where such neuroplastic changes would be expected. Prior studies have found changes to the power spectrum with aging (Chiang et al., 2011; Knyazeva et al., 2018; van Albada et al., 2010; Voytek et al., 2015), for example, age-related reductions in the frequencies of the alpha and beta band peaks. Voytek et al. (2015) suggest that such changes might indicate a change in the 1/f baseline possibly due to increased physiological noise with aging (also see Demuru & Fraschini, 2020). Furthermore, systematic longitudinal changes in the power spectrum have been observed following stroke (Giaquinto et al., 1994; Saes et al., 2020). Thus, implementing change classification to support longitudinal RS in clinical populations is an important future priority.