Reading is a remarkable human skill, yet the neural mechanisms underlying fluent reading remain incompletely understood. Parafoveal vision (the area surrounding the fixation point), despite its lower acuity, significantly contributes to reading fluency. Studies have shown that masking the parafoveal area severely impairs reading, highlighting its importance. However, there is considerable debate regarding the extent and nature of parafoveal processing. Models of reading differ on whether lexical information (related to word frequency) is processed in the parafovea before the eyes fixate on the word. Serial attention shift models propose that lexical processing is sequential, focusing on one word at a time. However, attention shifts to the next word before the eyes move, allowing for some parafoveal processing. In contrast, parallel graded processing models suggest simultaneous processing of multiple words within the perceptual span, with varying levels of attention allocated to each. Eye-tracking studies, while informative, indirectly measure parafoveal processing, often finding no impact of upcoming word frequency on fixation durations (parafoveal-on-foveal effect). Some studies using gaze-contingent boundary paradigms have revealed delayed effects, indicating influence from words beyond the next. This study employed the rapid invisible frequency tagging (RIFT) technique, combined with MEG and eye-tracking, to directly measure neural activity associated with parafoveal lexical processing during natural reading. RIFT, by flickering stimuli at high invisible frequencies, measures neuronal excitability related to attention. Previous research has shown RIFT successfully detects covert attention. This study aimed to determine whether lexical information is accessed from upcoming parafoveal words in natural reading by flickering target words subliminally during reading and measuring the resulting brain activity.

Existing research on parafoveal processing in reading is divided between serial and parallel models. Serial models, like the E-Z Reader model, posit that attention is focused on one word at a time, although attention may shift to the next word before the eyes do. This can explain phenomena like word skipping. Parallel models, such as SWIFT and the OB1-reader, propose graded allocation of attention across multiple words within the perceptual span, allowing for simultaneous lexical processing of both foveal and parafoveal words. Most eye-tracking studies support serial models, as fixation durations are typically unaffected by the lexical frequency of upcoming parafoveal words. However, studies using the gaze-contingent boundary paradigm have revealed delayed parafoveal-on-foveal effects, suggesting influence from words further ahead. The lack of consistent parafoveal-on-foveal effects in many studies has been interpreted as evidence against parallel models. However, the indirect nature of eye-tracking measurements motivates a more direct neural approach to examine this issue.

Thirty-nine participants (25 females) with normal or corrected-to-normal vision read 228 sentences silently while MEG and eye-tracking data were recorded. Sentences contained target words of either low or high lexical frequency. Target words were subliminally flickered at 60 Hz using RIFT throughout sentence presentation. The MEG data measured neural responses to this flicker while participants fixated on the pre-target word, allowing investigation of parafoveal processing without disrupting reading. Word length was matched between pre-target and target words across frequency conditions. To ensure comprehension, one-quarter of the sentences were followed by yes/no comprehension questions. Data analysis included time-resolved coherence between the 60 Hz flicker and MEG activity. Only sensors showing significant tagging responses (compared to a baseline) over the left visual cortex were included in subsequent analysis. Source localization using DICS confirmed the neural generators within early visual cortex. Pre-target fixation durations, target fixation durations, and gaze durations were analyzed. Coherence at 60 Hz during pre-target and target fixations was compared for low and high-frequency target words. The influence of orthographic information was examined by controlling for bi/trigram frequency. Finally, correlations between the pre-target coherence difference and individual reading speed were explored using a Spearman rank correlation. Additionally, Fixation-Related Fields (FRFs) were analysed to search for evidence of lexical parafoveal effects. A cluster-based permutation test was used to assess the significance of differences in FRFs between low and high frequency target words.

Eye movement data showed no effect of target word lexical frequency on pre-target first fixation durations, consistent with previous studies. Target fixation durations were longer for low-frequency than high-frequency words, confirming the manipulation of lexical frequency. MEG data revealed stronger 60 Hz coherence during pre-target fixations when followed by low-frequency compared to high-frequency targets. This effect was observed around 100 ms post fixation onset, suggesting early lexical parafoveal processing. Source analysis localized this activity to the left early visual cortex. No significant coherence difference was found during target fixations. Analyses controlling for orthographic factors did not alter the main findings. A positive correlation was found between the pre-target coherence difference (low minus high frequency) and individual reading speed, indicating faster reading speeds for participants showing stronger lexical parafoveal processing. Furthermore, analysis of fixation-related fields (FRFs) revealed significantly higher FRFs for pre-target words followed by low-frequency targets compared to high-frequency targets, further supporting the findings obtained via the RIFT method.

The findings challenge the serial attention shift models and support parallel graded models of reading. While eye-movement data did not reveal parafoveal-on-foveal effects, the neural evidence of early lexical parafoveal processing is not compatible with even the fastest serial models. This discrepancy might be because covert attention, reflected in neuronal activity, is not directly translated into overt attention (saccade initiation). The absence of a lexical parafoveal effect on pre-target fixation durations could even be beneficial for fluent reading; delaying the fixation on a difficult word in the low-acuity parafovea might be inefficient. The neuronal responses were found in early visual cortex, possibly due to interactive processing between higher-level lexical information and lower-level visual information. The stronger tagging responses for low-frequency words suggest allocation of greater covert attention to less familiar words, aiding processing. The lack of lexical frequency effect in foveal processing with RIFT could be attributed to differences in stimulus size, duration, and retinal receptor distribution, with parafoveal areas showing higher sensitivity to flicker. The results also show RIFT as a promising tool to investigate parafoveal processing and its relation to reading fluency and disorders like dyslexia.

This study provides compelling neural evidence for lexical parafoveal processing during natural reading, supporting parallel graded models of reading. The RIFT-MEG approach offers a powerful method to investigate the neural mechanisms of reading, complementing traditional eye-tracking and ERP methods. Future research could explore parafoveal processing at higher cognitive levels (semantic and syntactic) and its role in reading disorders.

The study’s limitations include the use of subliminal flickering, which might not fully reflect conscious parafoveal processing. Additionally, the specific neural mechanisms underlying the observed correlation between parafoveal processing and reading speed require further investigation. The use of a limited set of sentences also poses limitations.