1.1. Neural activity characterizing predictive processing

Predictive processing frameworks propose that perception emerges not merely from bottom-up sensory analysis, but through the continuous interplay between ascending sensory signals and descending predictions generated by internal models [2]. The Free Energy Principle (FEP) models the brain as a hierarchical inference engine that actively predicts incoming sensory data based on learned statistical regularities, continuously updating these predictions to minimize prediction error [1]. In this framework, perception (including perception of language) is fundamentally anticipatory: the brain generates expectations at multiple timescales, from milliseconds to seconds, and compares these predictions against actual sensory input. When predictions match the input (low prediction error), processing is efficient and requires minimal computational resources. When predictions fail (high prediction error), the system updates its models through error-driven learning. An important feature of the model is that the balance between relying on top-down predictions versus bottom-up sensory evidence is thought to shift with experience: as internal models become more refined through exposure to structured input, the brain can rely more heavily on prediction and less on detailed sensory analysis. Electroencephalography (EEG) can provide a window into predictive neural dynamics through the measurement of neural oscillations: fluctuations in electrical activity that reflect synchronized firing patterns across neural populations. The temporal structure of these oscillations is thought to organize information processing across multiple hierarchical levels [3,4]. Low-frequency oscillations (delta: 1–4 Hz, theta: 4–8 Hz) in particular have been implicated in both tracking and predicting slower temporal structures in sensory input, such as phrasal and sentence-level information in language [3]. The relationship between neural oscillations and stimulus structure can be quantified through coherence analysis, which measures the consistency of phase relationships between brain activity and external stimuli. Coherence, in this case, is a direct measure of neural entrainment: the degree to which brain rhythms synchronize with temporal regularities in the stimuli. From a predictive coding (FEP) perspective, enhanced coherence at specific frequencies suggests that the brain has learned to anticipate temporal structure at those timescales, generating oscillatory predictions that align with expected input patterns [1]. Research on spoken language comprehension has extensively documented such predictive entrainment. Neural oscillations in the delta and theta frequency ranges track the amplitude envelope of speech signals, with phase-locking to syllabic and prosodic rhythms facilitating word recognition and syntactic parsing [4]. This entrainment, however, is not merely stimulus-driven: it is modulated by linguistic knowledge, context, and predictability of the stimuli [5]. For instance, neural tracking of speech increases for intelligible versus unintelligible speech [6,7], and for predictable versus unpredictable linguistic sequences [8], suggesting that coherence reflects active prediction rather than passive following. Moreover, aging and experience affect these patterns: older adults show altered neural tracking of speech [9,10], and language proficiency modulates entrainment strength in second-language users [11,12]. These findings establish neural-stimulus coherence as a marker of experience-dependent predictive processing. Much less is known about how predictive processing operates in the visual modality, particularly for complex, structured visual input such as sign language (SL). While visual motion processing has been extensively studied using simple stimuli (e.g., drifting gratings, dot motion), naturalistic visual language presents unique challenges and opportunities for understanding prediction. Unlike auditory speech, where acoustic measures provide a simple temporal structure for entrainment analysis, sign languages encode linguistic information through multidimensional kinematic patterns, which include hand and face shapes and position, velocity and acceleration patterns, 3D spatial trajectories, and temporal coordination across multiple articulators (two hands, face, body). In sign languages, these motion dynamics carry phonological distinctions (e.g., straight vs. circular paths), morphological information (e.g., aspectual marking through movement repetition or speed modulation), and prosodic structure (e.g., phrase boundaries marked by holds and decelerations) [13,14]. The temporal organization of these features operates across multiple timescales simultaneously: from rapid articulator movements (8–15 Hz) to syllable-like sign units (4–5 Hz) to phrasal and sentence-level structures (0.5-2 Hz) [15–17]. This hierarchical temporal organization makes sign languages a rich testbed for examining multi-scale predictive processing in the visual domain.

1.2. Visual language and predictive processing

Sign languages present a unique opportunity to examine how lifelong linguistic experience shapes predictive mechanisms in the visual modality, independent of auditory processing. Deaf individuals who acquire sign language as their first language develop highly specialized neural systems for processing structured visual motion. Neuroimaging studies have shown that sign language comprehension recruits a left-lateralized fronto-temporal network similar to that engaged during spoken language processing [18–20], despite the difference in sensory modality. This suggests that linguistic computation drives core language network organization. Research using biological motion has demonstrated that Deaf signers process hand and body movements differently from non-signers, showing enhanced sensitivity to linguistically-relevant kinematic parameters and distinct patterns of neural activation in posterior temporal and parietal cortices [21,22]. These regions are thought to support motion prediction and temporal sequence processing, suggesting experience-dependent reorganization of visual prediction systems. Recent work has begun to examine neural entrainment to sign language using EEG coherence analysis. Our prior research demonstrated that Deaf signers exhibit frequency-specific neural tracking of optical flow (a measure of visual motion dynamics) in sign language videos [23,24]. Specifically, coherence between EEG signals and the temporal profile of visual motion was strongest in low-frequency bands, and localized to frontal and fronto-central scalp regions associated with language processing. This coherence pattern was specific to linguistically structured motion; when signers are presented with non-linguistic stimuli, the coherence pattern is substantially reduced or absent. This suggests that neural entrainment reflects not just frequency-following response in the visual domain, but predictive engagement with the temporal structure of linguistic units. No studies have yet examined how such predictive dynamics evolve across the lifespan in experienced sign language users, or how long-term experience shapes predictive neural mechanisms in the visual domain.

1.3. Aging effects on predictive processing

Aging introduces systematic changes to neural structure and function that have important implications for predictive processing. One well-documented phenomenon is the alteration of EEG spectral properties with advancing age, particularly an increase in spectral flatness—the flattening of the power spectrum such that power becomes more evenly distributed across frequencies [25] This flattening reflects what has been termed “neural noise”: increased randomness in neural firing patterns, reduced synchronization between neural populations, and weakened phase-amplitude coupling [26,27]. From an information-processing perspective, increased neural noise degrades neural representations and reduces the efficiency of neural communication. In the context of predictive coding, this creates a challenge: noisy neural signals make it difficult to maintain precise temporal predictions. Older adults, then, may struggle to generate or maintain accurate predictions about sensory input, particularly when the input has complex temporal structure. However, aging does not affect all cognitive systems uniformly. While structural measures such as spectral flatness show consistent age-related declines, functional measures of task performance often remain stable or even improve in domains where individuals have extensive experience [28]. This suggests that lifelong learning can compensate for age-related increase in neural noise, at least partially [10]. In the domain of language, older adults show preserved or enhanced use of contextual information and world knowledge to facilitate comprehension, even as lower-level perceptual abilities decline [29,30]. From a predictive processing perspective, this pattern is consistent with increased reliance on top-down predictions (built through decades of experience) to compensate for degraded bottom-up sensory signals. For Deaf signers with extensive, long-term use of sign language as their primary mode of communication, we might therefore expect a complex interaction between age-related neural changes (e.g., increased spectral flatness) and experience-dependent enhancement of visual-linguistic prediction. Specifically, while general spectral properties may show age-related decline, the specific neural mechanisms that track and predict sign language structure may be preserved or even enhanced, reflecting cumulative linguistic experience.

The present study investigates how lifelong visual language experience and age-related neural changes affect predictive processing in the visual-linguistic domain. We recorded EEG from Deaf users of Austrian Sign Language (ÖGS) spanning a wide age range (28–68 years) while they viewed videos of natural sign language sentences and time-reversed control stimuli. Time-reversed videos serve as a critical control: they contain identical visual motion statistics (same amount of movement, same velocity distributions) but lack interpretable linguistic structure. By comparing neural responses to forward versus reversed signing, we can isolate effects that are specific to linguistic prediction rather than general visual motion processing. We analyzed both the spectral properties of EEG, which provide information about general neural signal organization, and the coherence between EEG and visual motion (optical flow) in the stimuli, which provides measures of neural entrainment to temporal structure [23,31]. We pose three specific research questions:

1. 1) How does the strength and timing of neural entrainment to visual motion vary depending on whether the stimulus is linguistically structured (sign language) or not (time-reversed videos)?
2. 2) How are these coherence patterns localized in the brain—particularly between frontal and posterior regions known to support top-down and bottom-up processing, respectively?
3. 3) How do these coherence features change with age, reflecting both structural decline and the cumulative influence of lifelong language experience?

If visual language experience refines predictive mechanisms, we would expect age to be associated with increased coherence specifically during sign language processing (reflecting better-tuned predictions), potentially in posterior regions associated with visual motion analysis. Alternatively, we might observe age-related increases in response lag (time-shift) during processing of non-linguistic stimuli, reflecting difficulty in generating predictions for unpredictable input. By examining both general spectral features (which capture structural neural changes) and stimulus-specific coherence patterns (which capture functional predictive engagement), we can dissociate age-related decline from experience-dependent changes in visual linguistic prediction.

2.1. Participants

Individuals proficient in Austrian Sign Language (ÖGS) were recruited for the study. All participants reported normal or corrected-to-normal vision and had no history of neurological disorders. Their language proficiency was assessed through conversational interviews conducted by a certified Austrian Sign Language interpreter with 10 years of professional experience. This represents the gold standard for proficiency assessment in this linguistic community, as standardized psychometric tests for ÖGS proficiency do not currently exist. Participants varied in their age of acquisition of ÖGS; the majority were deaf from birth or early childhood, while a small number acquired ÖGS later in life. All participants were verified as proficient, active members of Austrian Deaf community through conversational interview with a certified interpreter. Because our analyses treat age as a proxy for cumulative linguistic experience, the relevant criterion is long-term, active use of ÖGS as a primary language; this criterion is met by all participants.

The study included 24 participants (13 male), aged between 28 and 68 years (M = 42, SD = 12.27). All procedures were approved by the University of Salzburg Ethics Commission (Protocol EK-GZ: 07/2015). Recruitment began on November 1, 2015, and ended May 31, 2016; Written informed consent was obtained from all participants prior to participation.

2.2. Stimuli and procedure

Each participant was presented with a set of video stimuli comprising 40 sentences signed fluently in ÖGS and their corresponding time-reversed versions, which were not linguistically interpretable. Additionally, 200 filler videos were included, consisting of sentences with classifier constructions, topicalized sentences in Subject-Object-Verb (SOV) and Object-Subject-Verb (OSV) word orders, and simple SOV-structured sentences. Neural and behavioral responses to filler videos, aimed at preventing habituation, are discussed in detail in [32,33]. Behavioral responses to fillers paralleled those for time-forward videos but were not included in the primary analysis due to differing spectro-temporal properties.

The sign language stimuli were presented as dynamic videos of signed sentences comprehensible to fluent signers (see Fig 1). A list of glossed sentence translations is provided in the appendix. To generate non-linguistic stimuli while maintaining identical spatiotemporal properties, videos were time-reversed. Participants rated the acceptability of each stimulus, including reversed videos, on a 7-point scale (1 = “not ÖGS,” 7 = “good ÖGS,” with 4 indicating “not ÖGS, but understandable”).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Analysis pipeline.

EEG recordings time-aligned to sign language and time-reversed video stimuli were processed along two parallel streams: (top) spectral feature analysis submitted to linear mixed-effects models examining effects of stimulus type and age; and (bottom) coherence between EEG and optical flow of the stimuli, with amplitude and timing features selected using ML pipeline (UFS → RFE) prior to age correlation analyses.

https://doi.org/10.1371/journal.pone.0346695.g001

Stimulus conditions were pseudo-randomized, ensuring that no condition appeared more than twice consecutively. Two balanced pseudo-random sequences were used across participants. Prior to the experiment, participants completed a training session to familiarize themselves with the task. Videos were displayed on a 35.3 x 20 cm screen at a resolution of 1280 x 720 pixels. Participants were instructed to minimize movement during video presentations. Each trial began with a fixation cross (2000 ms), followed by a 200 ms blank screen, and then the stimulus video. After the video, a question mark appeared for 3000 ms, prompting participants to rate the stimulus using a keyboard response.

2.3. EEG data acquisition and processing

EEG data were collected using a 26-channel system at a sampling rate of 500 Hz with active electrodes positioned according to the 10/20 system (Fz, Cz, Pz, Oz, F3/4, F7/8, FC1/2, FC5/6, T7/8, C3/4, CP1/2, CP5/6, P3/7, P4/8, O1/2). Electrodes were secured with an elastic cap (Easy Cap, Herrsching-Breitbrunn, Germany), and impedance was maintained below 5 k. Eye movements were recorded via electrodes placed at the outer canthi (HEOG) and superior/inferior orbital ridges (VEOG). AFz served as the ground electrode, and all recordings were referenced to the left mastoid.

During each trial, numerical triggers synchronized stimulus onset with EEG recordings. Offline processing included re-referencing to the average of both mastoids, and ocular artifact correction using the Gratton and Coles method [34]. EEG data were segmented from stimulus onset to 5 seconds post-onset, ensuring the analysis focused on responses to ongoing video stimuli (see Fig 1). While scalp EEG reflects volume-conducted activity from distributed neural sources, the regional organization of our electrode groups permitted inference about broad functional-anatomical patterns.

3.1. Spectral analysis of EEG signals

Spectral analysis was conducted to characterize the statistical properties of the EEG signal in the frequency domain, specifically focusing on measures of spectral complexity and energy distribution, separately for each of the regions of interest: frontal, posterior, left, and right hemispheres. The anterior region included channels FC1, FC2, F3, F4, and Fz. The posterior region comprised P3, P4, P7, P8, and Pz. The left hemisphere was defined by FC5, C3, CP1, CP5, and T7, while the right hemisphere included FC6, C4, CP2, CP6, and T8. The analysis included the computation of spectral entropy, spectral centroid, spectral spread, and spectral flatness. EEG data, which included 5-second segments time-locked to stimuli onsets at 500 Hz recording frequency, yielded frequency resolution .2 to 250 Hz.

Spectral descriptors were computed in MATLAB (The MathWorks, Inc.) utilizing the spectralEntropy, spectralCentroid, spectralSpread, and spectralFlatness functions, applied to the power spectral density (PSD) of each EEG segment. The PSD was estimated using a rectangular window.

Spectral entropy (H) was computed as a measure of the signal’s spectral complexity, defined by:

(1)

where P(f_i) is the normalized power at frequency bin f_i, and N is the total number of frequency bins. Higher entropy values indicate a flatter spectral distribution, reflecting increased signal unpredictability.

The spectral centroid (C) was calculated as the weighted mean of the frequency components, providing an estimate of the spectral “center of mass,” according to:

(2)

Spectral spread (S) quantified the dispersion of the power spectrum around the spectral centroid, computed as:

(3)

Spectral flatness (F) was defined canonically, as the ratio of the geometric mean to the arithmetic mean of the power spectrum:

(4)

All spectral features were computed for each EEG epoch and electrode channel, and subsequently averaged across trials to yield participant-level indices of spectral properties. These measures were then used in statistical analyses to investigate age- and experience-related changes in neural signal complexity and organization.

3.2. Optical flow analysis and coherence computation

Optical flow, a computer vision technique, was used to quantify motion between consecutive video frames. This measure tracks signal variation over time by calculating velocity magnitude (in pixels per frame) based on edge contrast. Optical flow preserves both spatial and temporal information, producing a velocity profile for each frame.

Optical flow computations were performed using MATLAB’s vision toolbox, yielding velocity matrices equivalent in size to the video frames. Histograms of velocity distributions were generated for each frame, and total motion magnitude was computed by summing velocity amplitudes across bins, resulting in an optical flow time series.

Coherence between the optical flow time series and EEG data was assessed at frequencies from .02 Hz to 12.5 Hz (limited by the 25 fps video frame rate). Timeseries were filtered at each frequency using a second-order IIR bandpass filter. Canonical correlation analysis, implemented via MATLAB’s NoiseTools toolbox [38], was employed to compute coherence between EEG signals and optical flow. Peak correlation values and corresponding time shifts were extracted for each frequency, stimulus, and electrode site. Sensor data was averaged over regions of the scalp (Anterior, Posterior, Left, Right) for each condition and participant. Anterior region included electrodes F3, F4, Fz, FC1, and FC2; posterior region included electrodes P7, P8, P3, P4, Pz; left hemisphere region included electrodes C3, FC5, T7, CP1, CP5; right hemisphere region included electrode channels C4, FC6, T8, CP2, CP6. This procedure produced the matrix with 496 features (4 regions x 62 frequency bins for peak correlation, and 4 regions x 62 bins for timeshift).

3.3. Feature selection pipeline

The coherence analysis generated 496 features (4 regions × 62 frequencies × 2 measures). Rather than testing each feature independently, which would require extensive multiple-comparison correction and severely reduce statistical power—we used a machine learning–based feature selection approach to identify the subset of features most informative for distinguishing linguistic from non-linguistic processing. This procedure served two complementary goals: (1) dimensionality reduction by isolating features that contributed meaningfully to stimulus discrimination, and (2) neurobiological interpretability by focusing subsequent analyses on neurally relevant signal components rather than noise. Conceptually, this is equivalent to defining functionally derived regions of interest in fMRI studies prior to hypothesis testing.

To identify neural coherence features predictive of age, we implemented a three-step feature selection and model interpretation pipeline. This approach was applied to coherence metrics computed between EEG signals and optical flow of sign language stimuli, including both time-shift and amplitude features across brain regions, using Python and the scikit-learn library. A preliminary application of this pipeline focusing on coherence-based classification of linguistic versus non-linguistic stimuli is reported in [39].

3.3.4. Spectral analysis and modelling.

Spectral features (entropy, flatness, centroid, spread) were analyzed separately across four anatomically defined electrode regions. Each feature quantifies a distinct property of the EEG signal’s frequency structure rather than time-varying power, as would be captured in a spectrogram. For instance, spectral centroid reflects the frequency “center of mass” of the power spectrum (analogous to acoustic brightness), spectral spread indexes bandwidth around that centroid, spectral flatness distinguishes narrowband or tonal signals from broadband or noise-like ones, and spectral entropy measures the overall disorder or uniformity of the spectrum. These descriptors are complementary and non-redundant: they summarize different aspects of signal organization within the analysis window rather than testing the same hypothesis multiple times.

Separate models were fitted for each region and feature because spectral features capture neural activity differences emerging at distinct timescales (task-driven versus experience-dependent) and because region-specific effects are theoretically and anatomically meaningful. We fitted 16 separate linear mixed-effects models: one model for each spectral feature (4 features: Spectral Entropy, Spectral Centroid, Spectral Spread, Spectral Flatness), separately for each brain region (4 regions: Frontal, Posterior, Left hemisphere, Right hemisphere). Models were fitted using maximum likelihood (ML) estimation. Subject-specific random intercepts accounted for baseline differences in spectral properties across individuals. No random slopes were modeled. The model for each feature f in region r was specified as:

(5)

where f_ij is the spectral feature value for participant i on trial j, is the fixed intercept, through are fixed effects for Age, Stimulus, and their interaction, u_i is the random intercept for participant i, and is the residual error. Stimulus was treatment-coded with Reversed video as the reference level (coded 0), such that reflects the difference in each spectral feature during Sign Language viewing relative to Reversed video viewing.

A summary of significant effects observed in the 16 separate linear mixed-effects models is provided in Table 1; all results for each model, including non-significant effects, are provided in S1 Table. To confirm that the region-specific patterns reported were not artifacts of analyzing ROIs separately, four additional omnibus models were fitted — one per spectral feature — including ROI as a fixed factor alongside Age, Stimulus, and all two- and three-way interactions; results are reported in S2 Table.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Summary of significant spectral effects of Age and Stimulus across brain regions.

https://doi.org/10.1371/journal.pone.0346695.t001

4.1. Behavioral results

Behavioral ratings collected using a 7-point Likert scale (7 = ‘that is good ÖGS’; 4 = ‘not ÖGS, but understandable’; 1 = ‘that is not ÖGS’) demonstrated a pronounced distinction in participants’ perceptions of the stimuli. Only the time-direct sign language videos were judged to be linguistically acceptable, while the reversed videos were overwhelmingly rated as lacking any resemblance to meaningful sign language. This categorical difference is reflected in the mean ratings, with time-direct videos receiving high scores (M = 5.80, SD = 1.48), while reversed videos were rated near the bottom of the scale (M = 1.72, SD = 1.35; see Fig 2. A paired-samples t-test confirmed a significant difference in ratings between conditions (t(23) = 14.01, p < .001), underscoring the drastic perceptual separation between the two types of stimuli. Importantly, response times did not significantly differ between conditions (t(23) = −1.3, p > .2), with comparable latencies observed for time-direct (M = 883 ms, SD = 535 ms) and reversed videos (M = 925 ms, SD = 541 ms).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Behavioral results for sign language and time-reversed video conditions.

(A) Acceptability ratings (7-point Likert scale) were significantly higher for sign language than reversed videos. (B) Reaction times did not differ significantly between conditions.

https://doi.org/10.1371/journal.pone.0346695.g002

4.2. Spectral analysis results

4.3. Interim summary: Spectral data

Age Effects. Across scalp regions, age was consistently associated with increased spectral flatness, suggesting greater spectral uniformity in older participants. This trend—interpreted as a form of neural dedifferentiation—was observed in the frontal region ( = .0000277, p < .001), posterior region ( = .0000316, p < .001), left hemisphere ( = 3.222 × 10⁻⁵, p < .001), and right hemisphere ( = 2.830 × 10⁻⁵, p = .003).

Stimulus Effects. Robust effects of stimulus type were observed for spectral entropy, with consistently higher entropy during Sign Language relative to reversed input. This indicates increased multiscale (broadband) complexity of neural response to linguistically meaningful stimuli. Significant stimulus effects on entropy were found in the frontal cortex ( = –.016, p < .001), the posterior cortex ( = –.014, p < .001), left hemisphere ( = –.019, p = .016), and right hemisphere ( = –.016, p = .037). Additionally, spectral spread and centroid in the left hemisphere were modulated by stimulus type ( = –.436, p = .032; = –.683, p = .027), reinforcing the left hemisphere’s sensitivity to linguistic features.

The regional specificity of age-related patterns was further confirmed by significant Age × ROI interactions in omnibus models for Spectral Entropy, Centroid, and Spread (all p < .001; S2 Table), indicating that age-related trajectories differ substantially across scalp regions and validating the region-specific analytical approach [40].

The omnibus model for Spectral Flatness additionally revealed a significant Age × Stimulus interaction ( = 4.091, p = .043), such that the age-related increase in spectral flatness was attenuated during Sign Language relative to Reversed video viewing ( = −5.856 × 10⁻⁶). This distributed effect, detectable only at the whole-scalp level, suggests that structured linguistic input partially counteracts age-related spectral dedifferentiation, consistent with experience-dependent stabilization of neural signal organization.

4.4. Feature coherence analysis results

Correlation analyses revealed distinct age-related effects across experimental conditions. In the reversed video condition (Condition 0), frontal timeshift at 0.2 Hz exhibited a significant positive correlation with participant age (r = .52, p = .023; Fig 3). This finding indicates that older participants showed delayed neural response timing to non-linguistic visual motion input.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Frontal neural response latency increases with age during non-linguistic visual motion processing.

Scatter plot showing a significant positive correlation between participant age and frontal timeshift at 0.2 Hz in the time-reversed video condition (linear fit with 95% confidence interval shown).

https://doi.org/10.1371/journal.pone.0346695.g003

Conversely, in the Sign Language condition (Condition 1), coherence in the posterior cortex at 0.4 Hz was significantly positively correlated with age (r = .51, p = .025; Fig 4). This suggests that older participants demonstrated increased neural synchronization with structured linguistic visual motion, consistent with experience-driven predictive processing.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Posterior neural coherence to sign language increases with age.

Scatter plot showing a significant positive correlation between participant age and posterior EEG coherence at 0.4 Hz in the sign language condition (linear fit with 95% confidence interval shown).

https://doi.org/10.1371/journal.pone.0346695.g004

Figs 3 and 4 illustrate representative features at low frequencies, which, based on our prior work ([41]), reflect the temporal scale of predictive processing in visual language comprehension. In the reversed video condition, older participants exhibited longer frontal timeshifts at 0.2 Hz—indicating delayed entrainment to unpredictable, non-linguistic visual input. In contrast, in the sign language condition, coherence in posterior regions (notably including parietal cortex—implicated in motion tracking and visual feature selection; see [42]) increased with age. This pattern suggests that older signers exhibit enhanced entrainment to predictable visual structure in linguistic stimuli, consistent with experience-driven refinement of predictive coding mechanisms.