Our purpose here is to present a novel model architecture meticulously designed to align with the requirements of the task at hand, modeling temporal–spatial dynamics of brain function for a specific task. In conclusion, our model represents a significant step forward in computational neuroscience, offering a powerful tool for analyzing and interpreting brain connectivity dynamics directly from raw rs-fMRI data. As a preliminary version of this study, future work will include additional experiments on specific tasks (such as classifying patients with schizophrenia or Alzheimer’s, predicting cognitive ability). Its utility and effectiveness opens up great potential to gain deeper insights into how the human brain adapts its functional connectivity specific to the task.

Results indicated that most atlas-based models (AAL, Dose, and CK) outperformed the Cube-based model, emphasizing the importance of anatomical priors for guiding the extraction of meaningful regional features. This process assisted in the precise identification of disease-relevant brain regions, improving classification performance for MDD detection. In future work, we plan to extend the proposed framework in several directions. First, we aim to investigate multimodal learning by incorporating resting-state fMRI and clinical features with sMRI. This integration is expected to provide complementary information that improves feature representations, optimizing classification accuracy and robustness.

Recently, one investigation focusing on first-episode, drug-naïve BD patients (mean age ≈ 20 years) analyzed depressive, manic, and mixed phases separately, finding ALPS index reductions specifically in the mixed phase. Together with our findings in early BD-D, these results preliminarily indicate that glymphatic alterations are present in the early stages of BD. However, the specific mood phase in which they occur remains unclear. Although we did not include a mixed-phase subgroup, the use of the iALPS pipeline and the well-balanced BD-D and BD-M samples enabled a robust comparison between depressive and manic phases. In addition, our study extends this evidence by demonstrating that reductions in ALPS indices showed a right-hemispheric predominance.

Taken together, the ALPS index emerges as a promising biomarker for glymphatic impairment in early-stage BD. However, methodological heterogeneity and cohort differences across existing studies preclude definitive conclusions. Furthermore, while the observed right-hemispheric reduction aligns with prior evidence of right-sided vulnerability in BD-D, it is important to emphasize that this finding emerged at a trend level and requires further confirmation. Future longitudinal and age-stratified studies are needed to delineate the developmental trajectory of glymphatic function and its phase-specific feature in early BD. A large cohort of young adults strengthens our study, however, several limitations still warrant consideration.

Conclusions In summary, while no significant group differences in neurometabolite ratios were observed between BPD patients and HC, our multimodal approach emphasizes the potential of integrating MRSI with EEG to investigate the neurobiology of BPD. Future research employing multimodal approaches is essential to deepen our understanding of the neurobiological underpinnings of BPD. Edited by: Francesca Strappini, Foro Italico University of Rome, Italy Reviewed by: Mahan Shafie, University of Turin, Italy Ibrahim Aslan, University of Southampton, United Kingdom

In SSD, hallucination severity correlated positively with STG–Heschl’s gyrus and TP–insula connectivity whereas negative symptoms correlated negatively with STG–insula connectivity. These findings suggest that there are distinct differences in FC between patients with OSSO and patients with SSD, which supports the proposal that OSSO should be treated as a separate clinical syndrome with distinct neural connectomes. Future research may explore whether interventions targeting these altered connectivity patterns could help reduce the risk of progression from OSSO to SSD.

In addition, multiple hypoconnectivity pathways within the AN were a distinct feature of SSDs with current AHs. These connectivity features differentiate OSSO from SSD and may reflect distinct neural correlates underlying AHs. Future research should test whether interventions targeting these connectivity abnormalities can improve clinical outcomes.

In contrast, we found no associations with aloof personality. Our results suggest that, within a sample exposed to PNMS, amygdala and hippocampal structure and function contribute differently to two different autistic-like characteristics, with hippocampus-motor connectivity explaining variance in communication impairment, and with hippocampal volume, amygdala- and hippocampus- sensory connectivity sharing the common mechanism in rigid behaviors. Given these links between brain and autistic-like traits, future research should examine whether brain volumes and connectivity mediate associations between PNMS and autistic-like traits in young adulthood.

After residualization, the association with right CMA volume became non-significant, whereas the associations with hippocampal volume as well as amygdala and hippocampal functional connectivity remained significant, supporting the robustness of our findings. A further limitation is the absence of a control group with similar levels of BAP traits but without a prenatal stressor. Without such a comparison, it remains unclear whether the neural markers identified in this unique context can generalize to typical autism populations, potentially limiting their diagnostic or therapeutic relevance. Therefore, the results need to be replicated in non-PNMS-exposed autistic samples. Despite the limitations, some strengths are inherent in this study.

Another proposal is that the hippocampal contribution to social cognition may also underlie effective communication: the hippocampus has been found to track social information in the physical environment, make predictions about the action of others and support appropriate responses and decision-making. Compared to this relatively novel role for the hippocampus, frequently-reported brain regions closely related to social cognition (referred to as “social brain”) involve other regions such as the medial orbitofrontal cortex and the right insula. Given social impairment as the core symptomatology domain of autism, future investigation is needed to gain a comprehensive picture of the neural circuits underpinning social processing. Meanwhile, our results showing that the motor-related regions (i.e., the left SMA and the left putamen) are implicated in pragmatic language deficits are supported by their previously reported roles. First, apart from the traditional function in motor control, the SMA plays a critical role in processing of speech communication and language reception; for example, the use of inner speech mechanisms during language encoding, lexical disambiguation, syntax and prosody integration, and context-tracking.

A previous study showed that, in 23 adolescents with autism spectrum disorder, increased BLA volume was associated with high reciprocal social interaction scores on the Autism Diagnostic Interview-Revised. The inconsistency between this previous study and our study might be explained by differences in the samples (nonclinical participants with a range of autistic-like traits versus clinically-diagnosed autistic individuals). Nevertheless, both studies have the limitation of small sample size, and thus, future studies with larger sample are warranted to verify the role of the BLA in social cognition. Another possible explanation is that the BLA is heterogeneous complex composed of four different subnuclei (i.e., basal nucleus, lateral nucleus, accessary basal nucleus and paralaminar nucleus), and that the basal nucleus and lateral nucleus might have closer involvement in social information processing. Therefore, there is also a need to further explore correlations of the two subnuclei structure and function with social deficits. [...] Specifically, it has been proposed that distortion is a characteristic of human memory and that some memory errors can be adaptive for flexible behaviors; conversely, overly rigid, robust, and undistorted memory may restrict the modification of thinking and result in low relational binding and behavioral inflexibility. There is evidence from human research that less flexible and less relational retrieval and learning may sometimes cause restricted and repetitive behavior, or, conversely, that inflexible behaviors in autistic individuals may result in rigid and less connected memories. Therefore, the present finding provides a clue that hippocampal CA1, CA3, CA4 and DG are implicated in rigid memory intertwining rigid behavior, suggesting a promising research direction for future studies to advance our understanding of their distinct roles. Finally, across the results, it was somewhat surprising that a single brain region exhibited associations with different subdomains of BAP depending on its structure and function. For example, CA1 volume, but not its functional connectivity, was associated with rigid personality, whereas CA1 functional connectivity, but not its volume, was associated with pragmatic language deficits.

A second strength of this study is our examination of the three core broad autism phenotypes separately: aloof personality, pragmatic language deficits and rigid personality. A third strength is the investigation of specific amygdala nuclei (i.e., BLA and CMA) and hippocampal subfields (i.e., CA1, CA3, CA4 and DG), recognizing that the amygdala and hippocampus are composed of heterogenous structures that may be differentially associated with specific broad autism phenotypes. Although the BAP assessment and the functional MRI scanning occurred concurrently at age 19, the neural correlates observed in this non-clinical sample are potential biomarker candidates that require further development in future research. Further, since the children at ages 16 and 19 self-reported their BAP phenotypes and underwent structural and diffusion MRI scanning, this design enables us to examine longitudinal changes in relationships between three BAP subdomains and amygdala and hippocampal volume and structural connectivity from age 16 to age 19. In conclusion, our results suggest that hippocampus-motor connectivity, which is responsible for emotion, memory and motor processing, possibly contributes to pragmatic language as seen in young adults.

Finally, although our in-house datasets provided detailed phenotypic information on the psychopathological, cognitive, and sensorimotor domains of SSD patients, this level of granularity is lacking in publicly available datasets. A validation using an independent, deeply phenotyped external cohort would have strengthened the robustness of our findings. Because cognitive testing and NSS were not performed in the HC group, it remains unclear whether the cognitive–motor profile observed in SSD patients is also present in healthy individuals. Future research efforts should focus on expanding and integrating well-characterized SSD cohorts, enabling more comprehensive analyses of the interplay between brain structure, cognition, and sensorimotor function in the disorder. Conclusion This multi-site Identified key predictive features, including GMV changes and ventricle enlargement, that shed light on the SSD’s pathophysiology and link structural deviations to functional outcomes, particularly in cognition and sensorimotor performance.

The leave-site-out CV approach validated the tool’s capacity for generalization, achieving results comparable to prior smaller-scale studies. Importantly, the identification of key age-dependent neuroanatomical features, such as total GMV and superior frontal sulcus, provided meaningful insights into the structural alterations in SSD. Additionally, the utilization of simple ML models, i.e., RF with almost standard hyperparameter settings further highlights a potential replication in future studies. The multivariate analysis further enhanced the understanding of brain-behavior relationships, offering a more integrated perspective on the disorder. Lastly, this is the first study that investigates the NSS and normative deviations of brain structures in different cohorts of SSD patients.

This study showed that up to 18% of SSD patients exhibited abnormally small mediodorsal and pulvinar thalamic volumes, and the extent of these deviations, rather than raw volumes, correlated with the severity of cognitive impairment. Notably, our findings are in accordance with previous studies, as we also observed that cognitive impairments in terms of higher TMT-B, DSST and CF scores in SSD are linked to alterations in the left inferior lateral ventricle and the left precentral gyrus. These associations remained robust under different analysis variations, shown in our sensitivity analysis, highlighting a potentially robust link for future research towards biomarker discovery. CT alterations in the postcentral sulcus were associated with higher PANSS negative, general and total scores, suggesting a potential link between somatosensory cortical morphology and SSD symptom burden. However, these relationships appeared sensitive to influential observations, most notably for PANSS general and total, so they should be interpreted cautiously and prioritized for replication in independent samples. [...] A validation using an independent, deeply phenotyped external cohort would have strengthened the robustness of our findings. Because cognitive testing and NSS were not performed in the HC group, it remains unclear whether the cognitive–motor profile observed in SSD patients is also present in healthy individuals. Future research efforts should focus on expanding and integrating well-characterized SSD cohorts, enabling more comprehensive analyses of the interplay between brain structure, cognition, and sensorimotor function in the disorder. Conclusion This multi-site Identified key predictive features, including GMV changes and ventricle enlargement, that shed light on the SSD’s pathophysiology and link structural deviations to functional outcomes, particularly in cognition and sensorimotor performance. Despite modest classification accuracy and considerable heterogeneity across individuals, these findings highlight the merit of NM in capturing interindividual variability—offering a more nuanced understanding of structural pathology.

Future research should integrate scanpath-based analyses to more comprehensively examine the temporal dynamics of visual search strategies in elderly users, potentially revealing nuanced differences in how older adults process complex interfaces. Additionally, while ethical approval was obtained, cultural variations in color perception (e.g., red’s significance in China) warrant cross-cultural comparisons. Future directions include comparing elderly with younger groups to delineate age-specific effects, testing additional moderators like blue fonts (as suggested in prior readability research), and developing adaptive interfaces using AI to dynamically adjust complexity based on user cognitive profiles, ultimately informing inclusive design for aging populations. This study investigates how the design of mobile news app screens affects how easy or difficult they are for older adults to use. We found that cluttered screens with a lot of information and mixed layouts make it harder for older users to find what they need, causing more mental effort and taking longer to complete tasks.

For older users, font color emphasis can significantly reduce cognitive load in interface search tasks. While this study provides valuable insights into how visual complexity and keyword color in mobile news interfaces affect elderly users’ cognitive load, several limitations should be acknowledged to guide future research. First, the sample was limited to 24 urban elderly participants in China with basic app familiarity, potentially restricting generalizability to rural or less tech-savvy seniors; future studies could employ larger, stratified samples across demographics to enhance representativeness. Second, the focus on visual complexity (e.g., layout and item count) and a single moderator (keyword color: red vs. black) overlooked other factors like font size, boldness, or interface novelty, which may interact with cognitive aging; expanding to these elements, perhaps integrating theories like selective attention models, would offer a more comprehensive framework.

In addition, the color of title keywords has a moderating effect on the influence of visual complexity of interface design on cognitive load of elderly users. For older users, font color emphasis can significantly reduce cognitive load in interface search tasks. While this study provides valuable insights into how visual complexity and keyword color in mobile news interfaces affect elderly users’ cognitive load, several limitations should be acknowledged to guide future research. First, the sample was limited to 24 urban elderly participants in China with basic app familiarity, potentially restricting generalizability to rural or less tech-savvy seniors; future studies could employ larger, stratified samples across demographics to enhance representativeness. Second, the focus on visual complexity (e.g., layout and item count) and a single moderator (keyword color: red vs. [...] Furthermore, while this study effectively utilized global eye-tracking metrics such as pupil diameter, fixation count, and blink frequency to assess cognitive load, it did not incorporate detailed analyses of scanpaths or saccadic movement patterns. This focus on overall physiological and attentional measures limits the interpretation of spatiotemporal aspects of visual navigation during search tasks. Future research should integrate scanpath-based analyses to more comprehensively examine the temporal dynamics of visual search strategies in elderly users, potentially revealing nuanced differences in how older adults process complex interfaces. Additionally, while ethical approval was obtained, cultural variations in color perception (e.g., red’s significance in China) warrant cross-cultural comparisons. Future directions include comparing elderly with younger groups to delineate age-specific effects, testing additional moderators like blue fonts (as suggested in prior readability research), and developing adaptive interfaces using AI to dynamically adjust complexity based on user cognitive profiles, ultimately informing inclusive design for aging populations.

Finally, the similarity between imagery and part-based viewing may arise from the fact that we used complex scenes as visual stimuli. Previous studies showed that eye movements follow more systematic and repeated scanpaths under high memory workload. Future studies could investigate whether these part-based patterns persist with less complex visual stimuli. To conclude, the present study provides direct evidence that eye movements during mental imagery are not simply replayed from perception. Instead, they resemble the spatiotemporal dynamics of part-based viewing, suggesting that they support the construction of mental images piece by piece.

Altogether, the integration of machine learning and deep learning into scanpath analysis marks a significant methodological shift. While these approaches introduce new challenges related to data heterogeneity, computational cost, and interpretability, ongoing progress in generative modeling, adaptive learning, and synthetic data generation offers promising avenues for overcoming these limitations. Ultimately, one of the most promising future directions lies in the development of hybrid frameworks that combine the interpretability of symbolic, AoI-based methods with the representational power of continuous, data-driven models, thereby enabling both robust quantitative analysis and meaningful cognitive interpretation. Edited by: Maria Chiara Caschera, National Research Council (CNR), Italy Reviewed by: Antonio Sarasa-Cabezuelo, Complutense University of Madrid, Spain Sarah Goodwin, Monash University, Australia Wolfgang Zangemeister, Montreal General Hospital, Canada

Recommendations This study demonstrates that individuals in the art-education group and the non-art group follow different visual evaluation processes, as evidenced by eye-tracking data. The findings indicate that individuals in the art-education group exhibit more systematic attentional processing toward AOIs and sustain attention for longer when evaluating images. Future studies may integrate eye-tracking metrics from individuals in the art-education group and the non-art group with machine learning methods to develop classification models and to examine how visual evaluation strategies evolve across educational levels. Eye-tracking studies can be conducted with groups differing in age and experience to determine how visual materials may be integrated into educational processes. To examine visual learning and cognitive processes, eye-tracking data can be analyzed using machine learning methods to develop classification models.

Differences between individuals with and without art education can be investigated more comprehensively to understand better factors shaping artistic perception. Longitudinal analyses can be conducted across different educational levels to evaluate how visual attention processes evolve as experience is gained. Based on these results, the following recommendations are proposed for future research: Overall, this study indicates that eye tracking is an effective tool for identifying cognitive differences within educational contexts and for revealing how visual scanning strategies vary as a function of training level. Future research may extend these analyses by applying similar methods across domains and examining eye-tracking metrics in broader, more diverse samples. Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). [...] Longitudinal analyses can be conducted across different educational levels to evaluate how visual attention processes evolve as experience is gained. Based on these results, the following recommendations are proposed for future research: Overall, this study indicates that eye tracking is an effective tool for identifying cognitive differences within educational contexts and for revealing how visual scanning strategies vary as a function of training level. Future research may extend these analyses by applying similar methods across domains and examining eye-tracking metrics in broader, more diverse samples. Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Scene descriptions in our study did not tap into precise compositional processes, such as symmetric relationships (Hafri et al.,), thematic role assignment in agentive verbs (Koring, Mak & Reuland,), or ergative verbs, which emphasize the object role in an action (see Khatin‐Zadeh, Hu, Eskandari, Banaruee, Yanjiao, Farsani & He for an example of gestures). Furthermore, since production and comprehension are not distinct and likely share predictive forward models (e.g., Kempen, Olsthoorn & Sprenger; Pickering & Gambi), a visual world version of our study may be able to better elucidate the moment‐by‐moment online integration of overt attention and sentence understanding (Snedeker & Trueswell,), and thus examine whether the evidence of semantic dominance observed in our data still holds. Future research could, for example, develop experimental designs that elicit descriptions involving actions, agents, and patients across languages, yielding a more comprehensive understanding of the LoV underpinning event representation and its reflection in overt attention. Additionally, by systematically manipulating syntactic constructions in their complexity, it is plausible to map their cartographic distances as a function of gaze alignment. While we expect semantic similarity to remain dominant, this controlled setting could isolate structural distinctions linked to predictable gaze shifts, effectively testing the comprehension–production interface (Pickering & Gambi,). [...] Computationally, the current study did not directly investigate real‐time alignment between visual and linguistic processing; instead, we focused on the association between the two information streams, using sentence similarity as a predictor of scan pattern similarity. However, our reliance on global similarity metrics may overlook fine‐grained individual differences in attentional strategies. Future research could benefit from network‐based scan‐path analyses, where eye movements are modeled as transition networks to reveal structural metrics (e.g., centrality, density) capable of distinguishing processing styles (Ma, Liu, Clariana, Gu & Li,). Additionally, recent advances in deep semantic gaze embeddings (Castner, Kuebler, Scheiter, Richter, Eder, Hüttig, Keutel & Kasneci,) and personalized scan‐path prediction models using observer encoders (Chen, Jiang & Zhao,; Xue, Xu, Mondal, Le, Zelinsky, Hoai & Samaras,) may offer higher granularity in distinguishing between participants who rely on top‐down linguistic guidance versus those driven by bottom‐up visual saliency. In parallel, the past decade has seen rapid advances in representing visual information using formalisms previously applied to linguistic information (e.g., dependency grammars, Elliott & Keller), leading to new developments in multimodal modeling (Koh, Fried & Salakhutdinov,), with applications to tasks such as image captioning (Elliott & de Vries,), text‐to‐image generation (Rombach, Blattmann, Lorenz, Esser & Ommer,), or sign language recognition (Li, Duan, Fang, Gong & Jiang,) (see Wu, Gan, Chen, Wan & Philip for a review of state of the art).

However, these changes were physiologically negligible. In addition, end‐tidal carbon dioxide partial pressure increased marginally (+1.1 ± 1.5 mmHg) during normocapnic hyperventilation, likely due to suboptimal tidal volume regulation, possibly because participants were required to focus simultaneously on both the eye movement task and voluntary breathing control. Given that this elevation is very small, we believe its influence such as on ventilatory drive would be negligible; however, this remains unclear in our study. Despite the slightly elevated end‐tidal carbon dioxide partial pressure observed during normocapnic hyperventilation, middle cerebral artery mean blood velocity paradoxically decreased (−8.5 ± 8.4 cm/s; Fig. 2B and  C ). [...] Despite the slightly elevated end‐tidal carbon dioxide partial pressure observed during normocapnic hyperventilation, middle cerebral artery mean blood velocity paradoxically decreased (−8.5 ± 8.4 cm/s; Fig. 2B and  C ). Since mean arterial pressure, cardiac responses and percutaneous oxygen saturation, all of which can influence cerebral blood flow, remained unchanged during normocapnic hyperventilation compared to pre‐hyperventilation levels, the reason for the observed reduction in middle cerebral artery mean blood velocity remains unclear. In addition, whether, and to what extent, the slight reduction in middle cerebral artery mean blood velocity influenced the latency of anti‐saccades remains uncertain. Nonetheless, it is important to note that the other oculomotor variables remained unaffected by normocapnic hyperventilation, indicating that this modest decrease in middle cerebral artery mean blood velocity likely had minimal impact on the overall results.

To mitigate these potential risks, voluntary regulation of breathing to prevent hypocapnia appears to be an immediate and effective intervention. Alternatively, inhalation of carbon dioxide‐enriched air during episodes of hyperventilation may attenuate hypocapnia, thus reducing the above‐mentioned risks. Future studies are warranted to directly investigate these possibilities. Considerations End‐tidal carbon dioxide partial pressure and middle cerebral artery mean blood velocity decreased (−0.3 ± 0.4 mmHg and −1.7 ± 2.1 cm/s, respectively) over time in the spontaneous breathing condition (i.e., before versus during the breathing intervention), indicative of a time‐dependent modulation (Fig. 2B and  C ). [...] It might be that oxygen level in arterialized capillary blood samples from fingertip was influenced by hypocapnia‐related factors such as reduction in finger blood flow (Umeda et al.,). We did not directly measure brain activity, brain temperature and reginal differences in brain blood flow. Future studies incorporating neuroimaging techniques (e.g., fMRI, MRS and EEG) would elucidate the underlying neural and physiological mechanisms of hypocapnia‐induced alterations in oculomotor behaviour. Conclusion We show that hyperventilation‐induced hypocapnia impairs oculomotor function in healthy young adults as evidenced by reduced fixation and saccade frequency, prolonged fixation duration, and shortened scanpath length during free viewing. Additionally, both hypocapnia and hyperventilation per se prolong saccadic latency in the anti‐saccade task.

We sample visual scenes with short gaze fixations separated by saccades. While low-level transsaccadic integration is known, semantic integration across multiple fixations remains unclear. We hypothesized that the brain predicts semantic information from one fixation to the next, and therefore postulated a neural signal associated with semantic novelty for each saccade. Novelty was measured using a deep network on foveal vision.

While seemingly related to novelty, this existing surprise measures distance to the previous video frame and neighboring pixels, not relative to the previous fixation. It also used hand-crafted, low-level features, not higher-level semantic features tuned for prediction across fixations. Future studies may explore alternative measures of novelty that incorporate a sequence of preceding fixations using recurrent vision models. Our measure of novelty did not take peripheral vision into account, which is known to influence foveation targets. Psychophysics research shows that integration of current peripheral vision with upcoming foveal vision is limited to attended targets. [...] Neural recordings show neurons responsive to foveal vision respond, shortly before a saccade, to the stimulus already visible in peripheral vision at the location of the upcoming fixation. Evidence for semantic processing of peripheral vision also comes from scalp recordings with neural responses to low-level features of the target prior to the upcoming fixation. Future studies should explore alternative measures of novelty that incorporate peripheral vision. Finally, we controlled for correlations between novelty and several low-level visual features, including luminance, spectrum slope and optical flow. However, any number of features, at various levels of complexity, could be used as additional controls.

So why use this method of measuring novelty? Here, we point out the initial motivation: explore whether self-supervised tuning mechanisms that are effective in computer vision may play a role in biological vision. Future work may explore if intermediate representations of candidate deep-network better match population activity observed at particular stages of the primate visual processing streams, and linking timing to various levels of processing. There is extensive modeling work on static images during a single fixation. In our view, the current results provide tight constraints on future computational models of free-viewing of dynamic visual stimuli.

Finally, we highlight the recent application of the D, I, R classification as a metric for gaze comparisons in the context of dynamic scenes, in which scan-path similarity metrics fail. We propose the D, I, and R classification as a computationally simple yet powerful tool to classify spatiotemporal aspects of scene fixations in an object-based and intuitive manner and provide well-documented code to implement it. Future research may explore potential functional differences between D, I, and R fixations.

A potential explanation for this is that the interplay between Detection and Return fixations contextualizes objects within a scene. In shorter trials, the prioritization of understanding contextual object (inter-)relations may take precedence over Inspections as it can play an important role in more efficient scene comprehension (Murlidaran & Eckstein,). Future experiments can test this hypothesis directly, e.g., by testing changes in D, I, R proportions as a function of object-scene congruency and viewing time (cf. Võ & Henderson,). An alternative explanation could be that the increased pace of exploration, with Detections spread across more objects and fewer Inspections, may lead to incomplete processing and therefore require more Returns, akin to regressions in reading (Engbert et al.,; Vitu & McConkie,).

Other types of images with considerably sparser or denser distribution of visual objects may require adjusted spatial heuristics, which could be determined e.g., by labelling part of the stimuli or combining the D, I, R approach with the detection of region of interests via hidden Markov models (Coutrot et al.,). The present findings provide evidence for systematic and reliable differences between Detection, Inspection and Return fixations and point to potentially diverging functional roles. Future studies should test the functional significance of each type of fixation in targeted experiments, e.g., probing the effect of working memory load or individual capacity on the proportion of R fixations. Future research could also investigate how individual differences in gaze behaviour (e.g., Broda & de Haas,; de Haas et al.,; Linka & de Haas,) are reflected in D, I, and R fixations and how they vary across age (Helo et al.,; Krishna et al.,; Linka et al.,). Further, oculomotor metrics like saccadic velocity and acceleration may reveal further facets of D, I, and R fixations and the differences between them.

neutral) expressions (Schurgin et al.,). These findings suggest that D, I, and R fixations may serve distinct perceptual functions during natural viewing, depending on the semantic nature of the target object. Future research should test this hypothesis by systematically manipulating object semantics and measuring how D, I, and R proportions vary. This could involve controlled experiments where object meaning, relevance, or contextual associations as well as low- and mid-level features are altered independently within the same stimuli. Furthermore, we found that specific predictions of D, I, and R fixations outperform the performance of a generic model including all fixations. [...] The current data show that this exploration is predominantly driven by high-level features, while Inspection fixations—successive fixations on the same object—are more dependent on geometric mid-level features. This suggests that semantic features are particularly important for the salience of extrafoveal targets, whereas subsequent fixations on a currently foveated object more strongly depend on geometric properties. Future research should use the D, I, R framework to explore such differences in feature weights in more depth to further test functional implications of D, I, and R fixations. Past research suggests a shift from a global to a focal processing mode across viewing time, indicated by a decrease in saccadic amplitudes and increase of fixation durations (Ito et al.,; Pannasch et al.,; Unema et al.,). Our results corroborate and extend these findings by showing an increase in fixation duration across D, I, and R as well as a descending frequency of Detection fixations and an ascending but quickly plateauing frequency of Inspection fixations. [...] is driven by Detection fixations and does not hold for Return fixations. Return fixations on the contrary seem to be restricted to close-by objects early in a trial and the corresponding saccades triple their amplitudes over the course of the first two seconds. Future research should use the D, I, R scheme to determine in how far this change is due to dynamic changes in (relative) Return saliency of peripheral objects, versus an effect of possible Return locations building up with exploration. Finally, we found shorter trial durations increase the proportion of Detections and decreases that of Inspections. This finding is in line with a hypothesized faster shift from exploitation to exploration when the overall viewing time is restricted. [...] An alternative explanation could be that the increased pace of exploration, with Detections spread across more objects and fewer Inspections, may lead to incomplete processing and therefore require more Returns, akin to regressions in reading (Engbert et al.,; Vitu & McConkie,). By providing an easily accessible and ready-to-use algorithm for the D-I-R classification we hope this classification approach will be adopted and prove useful for researchers from diverse fields, such as cognitive science, neuroscience, basic and applied vision research. For instance, future research could investigate whether Detection, Inspection and Return fixations and saccades rely on partially distinct neural processes. Some have suggested separate neural streams to be associated with distinct visual processing modes; the dorsal stream favouring a global processing of the visual space appropriate for exploration, and the ventral stream, responsible for a more central inspection of objects appropriate for exploitation (Sheth & Young,; see also Unema et al.,). To identify D, I, and R fixations, we used pixel masks of objects in the images. [...] The present findings provide evidence for systematic and reliable differences between Detection, Inspection and Return fixations and point to potentially diverging functional roles. Future studies should test the functional significance of each type of fixation in targeted experiments, e.g., probing the effect of working memory load or individual capacity on the proportion of R fixations. Future research could also investigate how individual differences in gaze behaviour (e.g., Broda & de Haas,; de Haas et al.,; Linka & de Haas,) are reflected in D, I, and R fixations and how they vary across age (Helo et al.,; Krishna et al.,; Linka et al.,). Further, oculomotor metrics like saccadic velocity and acceleration may reveal further facets of D, I, and R fixations and the differences between them. Finally, the D, I, R schema introduced here shows an interesting parallel to the well-established fixation taxonomy in reading research. [...] Finally, the D, I, R schema introduced here shows an interesting parallel to the well-established fixation taxonomy in reading research. First fixations on a word may be compared to object Detection fixations, refixations of a word to Inspection fixations and regressions to previous words in a text to Return fixations (Engbert et al.,; Vitu & McConkie,). While dominant models of eye movements during reading emphasize the role of lexical and linguistic factors to explain e.g., regressions (Engbert et al.,; Reichle et al.,), future research may probe whether there is mechanistic overlap between D, I, R patterns during scene viewing and corresponding viewing patterns during reading. Taken together, we find that the D, I, and R classification algorithm captures systematic and reliable differences between subgroups of fixations with diverging salience and dynamic profiles, suggesting they may map onto distinct functional classes of fixations. The approach has already been successfully applied to dynamic stimuli and for static stimuli without annotations, the D, I, and R classification can be approximated in a content-agnostic manner with a simple spatial heuristic.

The retrospective public datasets on which all analyses are based have variable annotation quality, small sample sizes for certain tasks (particularly survival prediction), and no paired endoscopy-CT data at the subject level, which limits the evaluation of true cross-modal complementarity. External validation is limited to a small number of centres and acquisition protocols, and comparisons with prior survival models rely on results reported in separate cohorts rather than direct head-to-head evaluation. Prospective, multi-center validation, the addition of other modalities like MRI, pathology, and molecular data, the creation of paired multimodal datasets to better analyze cross-modal interactions, and research into workflow integration and practical clinical impact are all important areas for future work.

We focused on the cis-effect variants and regulatory regions located within gene bodies or upstream and downstream of the gene. This allowed us to examine the transcription factor binding site-rich promoter and promoter proximal regions, as well as regions close to the 5′ and 3′ UTR, which have been shown to harbor regulatory motifs. Nevertheless, the more distal regions may still participate in gene transcription regulation through chromatin looping and will warrant future investigation using spatial regulatory information. In addition to CHD, we recognize that other developmental disorders such as specific subtypes of epilepsy, autism, attention-deficit/hyperactivity disorder, and neurocognitive disorders are thought to also be attributable in part to mutations in many genes that are developmentally regulated by thousands of regulatory elements. As the information about gene regulation accumulates, we anticipate that the approaches presented will be able to aggregate them at scale, thus enabling discoveries of disease-associated regulatory regions in developmental disorders that affect all human organs, including the brain.

We tested the effects of both single and short repeated doses (up to six days), but longer treatment durations are necessary to evaluate the longevity and clinical significance of HBK-15's cognitive benefits. While the present work centered on NMDA receptor hypofunction associated with psychiatric disorders, the anti-amnesic effects of HBK-15 warrant exploration in neurodegenerative settings. Future studies should establish whether its procognitive activity extends to models characterized by NMDA receptor overactivation and excitotoxicity. Addressing these limitations in future studies will enhance the translational relevance of our results. In conclusion, the multimodal compound HBK-15 emerges as a sigma-1 receptor agonist capable of restoring recognition and spatial memory under NMDA receptor hypofunction. [...] While the present work centered on NMDA receptor hypofunction associated with psychiatric disorders, the anti-amnesic effects of HBK-15 warrant exploration in neurodegenerative settings. Future studies should establish whether its procognitive activity extends to models characterized by NMDA receptor overactivation and excitotoxicity. Addressing these limitations in future studies will enhance the translational relevance of our results. In conclusion, the multimodal compound HBK-15 emerges as a sigma-1 receptor agonist capable of restoring recognition and spatial memory under NMDA receptor hypofunction. By rescuing hippocampal LTP and stabilizing oscillatory coordination, it engages both synaptic and network-level plasticity to counteract MK-801-induced impairments.

The identification and functional characterization of melatonin receptors in roots, as well as downstream signaling components, are critical for understanding melatonin perception and signaling specificity. In addition, post-transcriptional regulatory layers, including miRNA-mediated pathways and RNA modification, remain largely unexplored in the context of melatonin-regulated root development. Future studies integrating genetic mutants, cell-type-specific analyses, and multi-omics approaches will be essential to unravel the context-dependent roles of melatonin in root systems and to facilitate its rational application in sustainable crop production. Edited by: Andrzej Bajguz, University of Białystok, Poland Reviewed by: Krzysztof Mikołajczak, Polish Academy of Sciences, Poland Peihua Du, Tarim University, China Zhang Guoliang, Huaiyin Institute of Technology, China

At present, for HFD-induced obesity and metabolic diseases, there are strategies such as dietary regulation, supplementation of probiotics or prebiotics, FMT, MFH, etc., which effectively circumvent the pain points such as difficulty in adhering to traditional exercise and risks of drug treatment. Moreover, based on the research of high-altitude adaptive plants, combined with the dietary characteristics of high-altitude people, simulating the intestinal microbial ecosystem in a low-fat and high-fiber environment can further promote the synthesis of beneficial metabolites and repair the metabolic regulation association between organs. Future research needs to take the high-altitude animal intestinal adaptation model as an example, combined with multi-omics and germ-free animal models, systematically analyze the host signaling pathways triggered by the key functional strains and their characteristic metabolites of high-altitude animals, clarify the causal chain of their anti-obesity phenotype, and promote the obesity prevention and control toward a precise era. Edited by: Huan Li, Lanzhou University, China Reviewed by: Deli Xu, Qufu Normal University, China Xingen Yang, Shanxi Agricultural University, China

Inclusion of tendon-derived cells, endothelial populations, and vascular elements will be important to improve physiological relevance. Fifth, no functional assessment of the constructs was conducted at the endpoint. In addition to the absence of histological evaluation for tendon-like matrix deposition, such as collagen type I or tenomodulin, mechanical testing and analysis of ECM organization will be required in future studies to determine whether multiaxial loading promotes construct-level maturation. In addition, integrating sensors to monitor key nutrients, metabolites, and environmental factors such as oxygen levels, along with developing computational models, will provide a deeper understanding of the cellular environment and contribute to more effective tendon repair strategies. While passive shear or torsional deformation may occur within the bioreactor chamber, these components were not quantified in the present study.

This innovative approach allows for the application of torsion, compressive, and tensile stresses through a non-load-bearing membrane. These advancements have led to significant improvements in cellular orientation and pronounced activation of mechanosensitive signaling pathways, particularly PI3K/Akt, compared to the uniaxial platform. Our approach has demonstrated promising results; however, several limitations must be addressed in future work. The present study was intentionally designed to interrogate early mechanotransductive and cellular responses to complex loading, rather than long-term tissue maturation or construct-level functionality. First, although the robotic platform is capable of 3 kinematic DOF, only a single DOF was employed in this study, which, while relevant to supraspinatus loading, does not fully reproduce the complex motion patterns present in daily and skilled activities. [...] The present study was intentionally designed to interrogate early mechanotransductive and cellular responses to complex loading, rather than long-term tissue maturation or construct-level functionality. First, although the robotic platform is capable of 3 kinematic DOF, only a single DOF was employed in this study, which, while relevant to supraspinatus loading, does not fully reproduce the complex motion patterns present in daily and skilled activities. Future work will incorporate additional joint rotations to more accurately simulate physiological loading. Second, periodic media replacement does not adequately mimic the dynamic nutrient and waste exchange seen in vivo. A continuous perfusion system will be integrated to better simulate the native tissue environment and address potential challenges like continuous media supply and hypoxia during long-term culture. [...] While passive shear or torsional deformation may occur within the bioreactor chamber, these components were not quantified in the present study. Finally, although the MEF sensor exhibits reliable performance, its measured strain does not represent the strain of the scaffold or embedded cells, as it is not perfectly attached to the matrix. Future work will address these limitations through full integration of the sensor and matrix.

Future directions Looking forward, several emerging trends are expected to significantly advance the application of MRI in early detection and personalized intervention for pediatric NDDs. Particularly, artificial intelligence (AI) and deep learning are anticipated to markedly enhance the sensitivity and scalability of neurodevelopmental MRI. Recent investigation demonstrates that convolutional neural networks, auto encoders, and generative adversarial networks applied to pediatric structural and functional MRI are already achieving high accuracy. [...] This enables precision medicine in a way that aligns neurobiological phenotypes with therapy. Emerging studies of neuroimaging biomarkers in ASD and ADHD suggest that this stratification is within scope, particularly with advances in trans diagnostic and dimensionally informed models. Future directions in MRI driven for NDDs are shown in Figure 2. Finally, the convergence of multimodal biomarkers, combining genomics, imaging, and behavioral data, represents a major frontier. Multimodal MRI studies of ASD already integrate diffusion, structural, and perfusion measures, but coupling these with genetic information and longitudinal behavioral profiles could dramatically improve early risk models.

Applied to a deeply characterized healthy cohort, PhenoMol successfully predicted elite physical performance and outperformed regression models lacking network-based dimensionality reduction. Designed to be versatile and generalizable, PhenoMol enables studies across small and large populations to predict wellness, performance, and disease outcomes. The software is openly available to support future research in health, disease, and performance optimization. PhenoMol integrates graph theory and biological knowledge to reduce multi-omic dimensionality, predict phenotypes, and reveal causal patterns. It outperforms conventional models that lack biological constraints and is openly available for health, performance, and disease research.

Acute stress triggers the release of stress hormones such as cortisol, increasing stress reactivity and aiding post-stress recovery. Rodent studies revealed that stress reactivity is modulated by the gut microbiota, and few interventional studies have provided evidence for an effect on human cortisol dynamics. However, it remains unclear whether stress reactivity is related to interindividual variations in gut microbial composition and to one's capacity to produce microbial metabolites such as short-chain fatty acids (SCFAs). To close this gap, we analyzed data from 74 healthy human adults who completed the study in the laboratory and were either exposed to a well-established, standardized intervention that induced acute stress or to a non-stressful control condition (n = 35/39 per stress/control group). Stool samples were obtained at baseline, and the gut microbiota were characterized through 16 S rRNA gene amplicon sequencing.

Details of the post-hoc power analysis still indicated adequate power to detect medium-to-large-sized effects (Supplementary Results 6). These effect sizes fall within the range typically reported in human studies linking the gut microbiota to behavioral and physiological variables. We recommend that future studies use larger and more evenly distributed participant samples and integrate relevant behavioral, microbial, and physiological markers of post-stress recovery and potential differences pertaining to biological sex. To conclude, we investigated the role of the gut microbiota in modulating acute stress in healthy adults. Higher microbial diversity was associated with higher (both cortisol and subjective) stress reactivity after acute stress.

First, this study conducted questionnaire surveys exclusively among high-tech enterprises in Shandong, Jiangsu, and Shanghai. Given the pronounced regional disparities among China’s high-tech enterprises, this limitation may constrain the generalizability and external validity of the findings. Future research should expand the sample scope to include companies in central and western regions, as well as those with different ownership structures—such as state-owned enterprises and private SMEs—to enhance the external validity and generalizability of the results. Second, this study employs a mainstream empirical management approach using questionnaires to collect data. The selection of enterprise management team members and technical experts as survey respondents introduces a degree of subjectivity. [...] Second, this study employs a mainstream empirical management approach using questionnaires to collect data. The selection of enterprise management team members and technical experts as survey respondents introduces a degree of subjectivity. Future research will consider utilizing third-party ESG ratings or sustainability reports to assess ESG management practices. Third, this study only included firm age and firm size as control variables. Future research will consider incorporating industry competition, government subsidies, senior management team characteristics, and regional digital infrastructure to mitigate the influence of these factors on the adoption and performance of digital technology innovation. [...] Future research will consider utilizing third-party ESG ratings or sustainability reports to assess ESG management practices. Third, this study only included firm age and firm size as control variables. Future research will consider incorporating industry competition, government subsidies, senior management team characteristics, and regional digital infrastructure to mitigate the influence of these factors on the adoption and performance of digital technology innovation. Fourth, further research may include ownership structure, industry competitiveness, and digital infrastructure as control variables to enhance the explanatory power of the model. Fifth, the intuitive presentation of the moderation effect diagram is somewhat lacking. [...] Conversely, in other sectors like technology and services, environmental concerns may not be prioritized, leading to insufficient investment in ESG management and consequently weaker moderation effects. Combining different types of firms in regression analysis can result in a visually diminished effect. Future research could group companies by industry to develop moderation effect diagrams, enhancing clarity and persuasiveness. Sixth, since the data for this study were obtained through questionnaire surveys, although necessary checks for reliability, validity, and common method bias were conducted to ensure the data fall within reasonable ranges, the selection of questionnaire items relies on subjective measurements by respondents, which may introduce some interference into the results. In future research, we will adopt a combined approach of panel data and case studies to ensure the scientific rigor and accuracy of our conclusions. [...] Future research could group companies by industry to develop moderation effect diagrams, enhancing clarity and persuasiveness. Sixth, since the data for this study were obtained through questionnaire surveys, although necessary checks for reliability, validity, and common method bias were conducted to ensure the data fall within reasonable ranges, the selection of questionnaire items relies on subjective measurements by respondents, which may introduce some interference into the results. In future research, we will adopt a combined approach of panel data and case studies to ensure the scientific rigor and accuracy of our conclusions. Seventh, the comprehensive and thorough investigation of these variables (digital technology innovation, organizational resilience, ESG management, and enterprise performance) may not be sufficient. In future research, the specific technical applications of digital technology innovation, the specific manifestations of organizational resilience, and the specific practices of ESG management can be expanded to further analyze the profound impact of these variables on enterprise performance. [...] In future research, we will adopt a combined approach of panel data and case studies to ensure the scientific rigor and accuracy of our conclusions. Seventh, the comprehensive and thorough investigation of these variables (digital technology innovation, organizational resilience, ESG management, and enterprise performance) may not be sufficient. In future research, the specific technical applications of digital technology innovation, the specific manifestations of organizational resilience, and the specific practices of ESG management can be expanded to further analyze the profound impact of these variables on enterprise performance. Eighth, this study only considered the influence of digital technology innovation and organizational resilience on enterprise performance. Looking ahead, strategic flexibility can be included as an intermediary variable in the research scope, and green innovation and digital transformation can be included as moderating variables in the scholarly work.

Third, this study only included firm age and firm size as control variables. Future research will consider incorporating industry competition, government subsidies, senior management team characteristics, and regional digital infrastructure to mitigate the influence of these factors on the adoption and performance of digital technology innovation. Fourth, further research may include ownership structure, industry competitiveness, and digital infrastructure as control variables to enhance the explanatory power of the model. Fifth, the intuitive presentation of the moderation effect diagram is somewhat lacking. From an industry-specific perspective, this may stem from the fact that in sectors closely tied to environmental issues—such as energy and manufacturing—ESG management plays a significant moderating role.

Overall, this study presents an integrated GeoAI-based wildfire susceptibility modelling framework that combines stacking ensemble learning with explainable artificial intelligence and uncertainty–sensitivity analysis. The proposed approach offers a reproducible methodology for identifying wildfire-prone areas and improves understanding of wildfire susceptibility patterns in complex Himalayan environments. Future research may further extend this framework by incorporating dynamic climate projections and socio-environmental scenarios to assess potential changes in wildfire susceptibility under evolving environmental conditions.

Conclusions and future research directions In this paper, a fault-tolerant VECC architecture with mobility awareness for vehicular task offloading was proposed. The proposed bi-level DQN architecture comprises a level-1 DQN agent for high-level scheduling and level-2 DQN agents operating at each RSU. The level-1 DQN agent determines the optimal RSU for task execution by considering real-time workload and network latency, while the level-2 DQN agents jointly decide the node assignment and select the most appropriate recovery pattern among First Result, Recovery Block, and Retry. [...] From a practical standpoint, the proposed model offers a promising solution for next-generation ITS and vehicular edge networks, where delay-sensitive and safety-critical applications, such as cooperative perception, autonomous driving, and real-time traffic coordination, require reliable and adaptive task execution. The ability to dynamically balance workload distribution and fault-tolerance strategies provides a foundation for resilient and efficient vehicular edge computing deployments. Future research will focus on extending the framework toward scalable and delay-tolerant learning architectures through asynchronous federated DQN mechanisms.

In this study, model reliability was ensured by restricting the analysis to clean, well-consolidated sandstone intervals, where the NPHI–Wyllie porosity estimation and the underlying petrophysical assumptions remain valid. It must be noted, however, that the approach has not yet been tested in mixed lithologies, carbonates, shales, or clay-rich units; thus, its sensitivity to these rock types cannot be evaluated with the current dataset. Future work should focus on extending, calibrating, and validating the methodology in more heterogeneous lithologies to assess its broader applicability and potential limitations. The methodology presented in this article has been applied exclusively to sandstones, primarily for two reasons. First, elastic measurement data from various sandstone types were available, which are essential for calibrating the elastic moduli response to seismic velocities.

Our findings highlight the urgent need for more robust collaboration protocols, adversarial-resistant debate frameworks, and principled guardrails governing LLM-to-LLM communication. With the growing deployment of LLM agents in coordinated and autonomous environments, designing techniques to mitigate persuasive manipulation will be increasingly relevant to ensuring the reliability and safety of multi-agent AI systems 48 . Also, future research should therefore prioritize structural and protocol-level defenses—such as cross-agent consistency analysis and verification-aware debate mechanisms—over purely prompt-based mitigation strategies.

The organization of the phase of electrical activity in the cortex is critical to inter-site communication, but the balance of this communication across large-scale (>8 cm), macroscopic (>1 cm), and mesoscopic (1 cm to 1 mm) ranges is an open question. The spatial frequencies (i.e. the spatial scales) of cortical waves have been characterized in the gray matter for micro- and mesoscopic scales of cortex and show decreasing spatial power with increasing spatial frequency.