Population aging is a major demographic trend [1]. Italy ranks second, behind Japan, and is among Europe’s longest-lived countries [2, 3]. This shift has serious consequences for healthcare systems. Advancing age raises the risk of multimorbidity, including neurocognitive disorders [4]. Alzheimer’s disease (AD), the leading cause of dementia, is marked by progressive cognitive, functional, and behavioral decline, and is a major cause of disability [5]. The latest World Alzheimer Report estimates that over 50 million people worldwide have dementia, a number expected to triple by 2050 [6]. Mild cognitive impairment (MCI), a transitional state between normal aging and dementia, is also common and may progress over time [7, 8]. Early identification of cognitive decline is crucial, as it enables timely intervention, preserves abilities, and improves the lives of older adults and caregivers [9].

This issue is especially important in nursing homes, where residents are often frail and medically complex, and at increased risk of cognitive deterioration [10]. Earlier recognition of cognitive decline in these settings can support more timely, personalized care planning and facilitate non-pharmacological interventions, such as cognitive stimulation and computerized training [11]. Early detection may also reduce adverse events related to unrecognized decline, wandering, falls, and agitation, which increase staff burden and care costs in residential facilities [12, 13]. Progression from MCI to dementia further raises care needs and supervision requirements, affecting staffing and resource allocation in nursing homes [14]. However, cognitive impairment in institutionalized older adults often goes unrecognized, and early detection can be harder than in traditional clinical contexts [15]. Thus, digital markers, including linguistic and eye-tracking metrics, have been proposed as promising, sensitive, non-invasive tools for detecting subtle cognitive changes in neurodegenerative disorders [16, 17].

Reading is a highly complex cognitive task that involves the coordinated integration of perceptual, linguistic, and executive processes [18]. Although reading paradigms are commonly used in eye-movement research, the existing evidence on aging and cognitive impairment remains varied and should be interpreted carefully [19, 20].

Previous studies suggest that reading aloud may remain relatively intact in some individuals with mild AD, even when comprehension is impaired. Single-case observations by Ripamonti et al. [21] support this dissociation. Martínez-Nicolás et al. [22] also found that AD patients could accurately read words without necessarily understanding their meanings. The same study indicated that detailed temporal measures of reading fluency, such as speed and pausing, might more sensitively distinguish early AD from asymptomatic controls than broader global assessments, such as the Mini-Mental State Examination (MMSE). Education level and cognitive reserve could also influence performance [23].

Evidence from eye-tracking studies is similarly mixed. A meta-analysis by Moreno et al. [24] showed that healthy older readers tend to have longer fixation durations, more regressions, and longer overall reading times than young adults, suggesting decreased efficiency despite preserved basic reading ability. In clinical populations, Lueck et al. [25] observed slower reading and altered oculomotor behavior in patients with probable mild-to-moderate AD, including longer fixations, more forward saccades, and more regressions. However, saccade duration was similar to that of healthy controls. More recent multimodal studies further suggest that lexical difficulty may significantly influence oculomotor behavior. For example, Shah et al. [26] reported that high-difficulty words elicited more atypical gaze patterns in AD than in MCI or control participants, whereas low-difficulty words provided less information. Similarly, Groznik et al. [27] found that cognitively impaired participants read more slowly and with more irregularity. Nonetheless, they concluded that a short reading task was better suited for a broader assessment battery than as a standalone test.

Overall, these findings indicate that measures derived from reading can provide valuable insights into study outcomes. However, substantial methodological differences across reading tasks, clinical classifications, oculomotor metrics, eye trackers, and data collection protocols limit direct comparisons between studies and hinder practical application. Evidence is particularly limited among nursing home populations, where ecological constraints, participant frailty, and concerns about data quality complicate feasibility and interpretation.

Reading aloud with comprehension is a complex and ecologically meaningful behavior. It depends on the integration of perceptual, lexical, phonological, semantic, attentional, and executive processes. Therefore, reading is a valuable task for detecting cognitive decline, as measured by accuracy, comprehension, and efficiency. Established cognitive models inform the present study of reading and comprehension. Dual-route word recognition models, such as the Dual Route Cascaded model [28], help interpret reading accuracy and error patterns through lexical and sublexical processing. Construction–integration accounts of comprehension, such as Kintsch’s model [29], support our understanding of comprehension performance in terms of semantic integration, working memory, and executive control. Models of eye movement control, such as E-Z Reader [30], provide a foundation for considering oculomotor indices as markers of online lexical and attentional processing. In this study, E-Z Reader is used solely as a theoretical reference for the eye-tracking component. A schematic appears in Supplementary Fig. 1. No model fitting or parameter estimation was performed. Instead, the study was conducted as part of the CogET project, which investigates language and cognition in nursing home residents. The approach combines neuropsychological assessments with a reading task supported by multimodal recordings, including audio and, when possible, eye-tracking. Reading-derived linguistic performance measures were selected as primary outcomes because of their low burden, ecological validity, and direct links to reading accuracy, comprehension, and fluency. Eye-tracking was added as an additional exploratory process measure.

The objectives of this cross-sectional pilot study were fourfold. First, we aimed to assess the feasibility of administering an ecological reading task in a nursing home, including task completion, audio recording acquisition, eye-tracking calibration and validation, and the proportion of usable recordings. Second, we aimed to compare the cognitively impaired (CI) and healthy control (HC) groups on primary reading outcomes, such as total reading errors (TRE), total comprehension errors (TCE), and words per minute (WPM), as well as secondary outcomes like Total Reading Time and specific error types. Third, we sought to evaluate how well reading-derived measures distinguish between CI and HC groups. Fourth, we aimed to examine the relationships between reading-derived measures and neuropsychological, functional, affective, and cognitive reserve measures. Given the feasibility-focused and exploratory nature of this pilot study, we did not predefine directional hypotheses for all outcomes. Instead, we focused on estimating feasibility parameters, between-group differences, discriminative ability, and covariate-adjusted associations. The results will help guide the design of future confirmatory studies.

The study followed STROBE checklist guidelines and is part of the CogET project, pre-registered on the Open Science Framework (OSF DOI: https://doi.org/10.17605/OSF.IO/VG3TD). Data collection occurred from June to December 2025. Recruitment used a convenience sampling strategy, considered appropriate for exploratory and pilot studies [31].

Participants were permanent residents of the Saint George nursing home in Cavallermaggiore, Cuneo, Italy. Inclusion criteria were: (a) age 65 years or older; (b) residence in the nursing home at the time of recruitment; and (c) ability to read a standard Italian text displayed on a computer screen. Exclusion criteria were: (a) a documented neurological or psychiatric disorder other than mild cognitive impairment or Alzheimer’s disease; (b) a history of epilepsy; (c) uncorrected visual impairment or ophthalmological conditions interfering with eye tracking; and (d) inability to provide informed consent. All participants had normal or corrected-to-normal vision and were able to understand and complete the experimental tasks.

A total of 60 participants were enrolled, comprising 30 healthy controls and 30 older adults with CI of varying severity. The group assignment was determined based on the overall neuropsychological profile and the clinical evaluation conducted by two qualified neuropsychologists. Healthy controls were required to have a MMSE score above 26, interpreted according to Italian normative values corrected for age and education, and no clinical evidence of cognitive impairment. Participants in the CI group showed clinical evidence of cognitive decline supported by neuropsychological assessment. The MMSE was used as a global cognitive screening measure, whereas the Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) and Frontal Assessment Battery (FAB) contributed to the broader characterization of the cognitive profile and supported clinical classification.

All participants underwent a neuropsychological, affective, and functional assessment.

Global cognitive functioning was screened using the MMSE [32]. The MMSE includes 30 items evaluating orientation, short-term memory, attention and calculation, learning, language, and constructive praxis. Scores were interpreted using the Italian standardization by Magni et al. [33], which provides normative adjustments for age and education. In this study, the MMSE served as a general cognitive screening tool, and healthy controls were required to score above 26.

A broader cognitive profile was obtained using the RBANS [34], a brief standardized battery comprising 12 subtests that yields a Total Scale score and five Index scores: Immediate Memory, Visuospatial/Constructional abilities, Language, Attention, and Delayed Memory. The RBANS provides age-adjusted standard scores, aiding in characterizing cognitive performance across domains and supporting comparisons between groups with varying levels of cognitive decline [35, 36]. In this study, the RBANS was used to support the clinical determination of the cognitive profile.

Executive functioning was further evaluated with the FAB [37], a six-subtest screening tool for frontal executive functions, including conceptualization, phonemic fluency, motor programming, sensitivity to interference, inhibitory control, and environmental autonomy. FAB scores were interpreted using the Italian norms by Appollonio et al. [38], adjusted for age and educational level.

Affective state was assessed using the Geriatric Depression Scale (GDS) [39], administered in its Italian version [40]. Functional independence was measured with the Barthel Index [41, 42]. Lastly, cognitive reserve was estimated using the Cognitive Reserve Index questionnaire (CRIq), a semi-structured interview that quantifies lifetime cognitive enrichment through education, occupational attainment, and cognitively stimulating leisure activities [43, 44]. The GDS, Barthel Index, and CRIq were used to characterize affective, functional, and reserve-related aspects of the sample, but were not criteria for group assignment.

Eye movements were recorded in a dedicated room within the nursing home, easily accessible to residents and free of architectural barriers. Recordings were obtained using a Gazepoint GP3 video-based eye tracker sampling at 60 Hz (Gazepoint Research Inc., Vancouver, BC, Canada). According to the manufacturer, the device has a spatial accuracy of approximately 0.5–1.0° of visual angle under optimal conditions and a spatial resolution of approximately 0.1°. Binocular gaze data were streamed through the Gazepoint Application Programming Interface (v7.1.0, Gazepoint Research Inc.).

The experiment was run on a laptop computer (Intel Core i5-1235U processor, 16 GB RAM, Windows 11) (Dell Inspiron 15, Dell Inc., Round Rock, TX, USA) using a custom PsychoPy script (Release 2023.1.3, Open Science Tools Ltd., Nottingham, UK) [45]. Stimuli were presented on a 24-inch Dell Full HD monitor (Dell Inc.) with a resolution of 1920 $\times$ 1080 pixels. The viewing distance was maintained at approximately 60–65 cm (mean = 62.45 cm) from the participants’ eyes, in accordance with the device recommendations. The Gazepoint monitoring window and pre-task positioning checks were used to optimize setup. Participants sat comfortably in front of the screen in a quiet, well-lit room. No chin rest was used, but participants were asked to minimize head movements during recording. The reading task setup and procedure are summarised in Fig. 1.

Fig. 1.

The reading paradigm procedure. (1) Experimental Setup and extracted features. (2) Description of the procedure developed in PsychoPy. (3) Questionnaire for the evaluation of comprehension. Created in BioRender (https://www.biorender.com/). Cecchetti, S. (2026) https://BioRender.com/2m4mv47.

The reading stimulus was the ‘sandwich’ story, a 177-word passage spread over 15 lines, selected from the Discourse Comprehension Test (DCT) included in the Narrative Comprehension and Production Test battery (NCPT) [46]. This text was chosen for multiple reasons. First, it offers a brief, meaningfully realistic reading task suitable for older adults in a nursing home setting. Second, it was presented in the participants’ native language. Third, it is part of a test battery with prior clinical use, including work with adults with AD [47]. Finally, following recommendations by Brookshire and Nicholas [48], the passage contains a mildly humorous element intended to enhance engagement and lower tension during testing. Participants were instructed to read the passage aloud as quickly and accurately as possible while also trying to understand its content, as comprehension questions would follow. Instructions first appeared on the screen and were then repeated orally by the examiner to ensure understanding of the task. Just before the text was presented, a fixation cross was shown to focus gaze on the screen. The maximum time allowed for the reading task was set at 180 seconds.

Comprehension was assessed immediately afterward using eight structured questions related to the passage (see Supplementary Table 1). These questions targeted different levels of processing, including explicitly stated information (MIS), implicit main ideas (MII), and peripheral details presented either explicitly (DTS) or implicitly (DTI). This structure allowed for characterizing performance not only in terms of literal information retrieval but also in relation to inferential processing and the integration of implicit and secondary story elements. The questions were administered in a paper-and-pencil format. The overall reading paradigm is illustrated in Fig. 1.

Eye-tracking pre-processing and analysis followed the general logic of the gaze analysis pipeline described by Duchowski [20], including raw data inspection, event extraction, area-of-interest assignment, and aggregation of summary measures for statistical analysis. Raw gaze data consisted of time-stamped horizontal and vertical gaze coordinates. From these data, fixation- and saccade-related measures were derived and summarised using custom Python scripts (Python 3.10, Python Software Foundation, Beaverton, OR, USA).

Specifically, the raw data stream (the raw gaze signal) output by the eye tracker consists of gaze points $g_i=(x_i,y_i,t_i)$, where $x_i$ and $y_i$ are the 2D gaze point coordinates on the screen and $t_i$ is the timestamp of every sample. A third-order $\left(s=3\right)$ Savitzky-Golay (SG) filter $h_i^{t,s}$ of width $p=5$ is used to differentiate the raw positional gaze signal into its velocity estimate:

$$\dot{x_n^s}(t)=1/\left(\mathrm{\Delta }{t}^s\right)\left(\sum _{i=-p}^p{h}_i^{t,s}{x}_{n-i}\right)$$

applied independently to the $x_i$ and $y_i$ 2D gaze point coordinates. The resulting velocity signal $(\dot{x_i},\dot{y_i})$ is thresholded at 36 degrees per second. The signal where velocity exceeds this threshold is labeled as part of a saccade; elsewhere, the signal is labeled a fixation. See Duchowski [20] for further details.

For text-based eye-tracking analysis, Areas of Interest (AOIs) were defined at the word level, with one AOI assigned to each word of the passage, following a word-based approach similar to that described by Busjahn and Tamm [49]. AOIs were created in document editing software (Scribus, v. 1.7.0, The Scribus Team, open-source, https://www.scribus.net/) based on the exact visual layout of the displayed text. Gaze events were then assigned to the predefined AOIs, and events falling outside these word-level AOIs were excluded from AOI-based analyses. A more detailed description of AOI construction and the custom processing workflow is provided in the Supplementary Fig. 2 to facilitate reproducibility.

The resulting eye-tracking metrics included conventional aggregate oculomotor measures, such as fixation count, fixation duration, and saccade amplitude, as well as AOI-based aggregate indices. Because the study was conducted in an ecological nursing home setting, without head stabilization and using a 60 Hz device, the analyses focused on robust summary measures rather than highly fine-grained temporal parsing.

Audio recordings were processed separately. Each file was manually reviewed in ELAN (version 7.0, Max Planck Institute for Psycholinguistics, Nijmegen, Netherlands) [50, 51], where reading errors, including mispronounced words, repetitions, excess words, and omissions, were identified and annotated.

The study outcomes were predefined as feasibility, primary, secondary, and exploratory outcomes. Feasibility outcomes included completion of the reading task, successful acquisition of audio recordings, successful eye-tracking calibration and validation when attempted, and the proportion of usable audio and eye-tracking recordings available for analysis.

The primary outcomes focused on key reading metrics: TRE, TCE, and WPM. TRE was defined as the total number of reading errors made during aloud reading, including substitutions, omissions, repetitions, excess responses, and mispronunciations. TCE was defined as the total number of incorrect responses on the comprehension questions administered after the reading passage. WPM was defined as the number of correctly read words per minute.

Secondary outcomes included Total Reading Time, individual components of reading and comprehension performance, and timing-based eye-tracking metrics. Reading error types comprised substitutions, omissions, repetitions, excess responses, and mispronounced words. Comprehension error types included MII, MIS, DTI, and DTS errors. Timing-based eye-tracking measures included eye-tracking-based WPM and Total Reading Time.

Exploratory outcomes encompassed traditional first-order eye-tracking metrics, such as fixation- and saccade-based indices. Since usable eye-tracking data were available only for a smaller subsample, analyses involving these measures were considered exploratory.

Descriptive statistics are reported for socio-demographic variables, neuropsychological measures, and reading-derived outcomes, stratified by group (HC vs CI). For between-group comparisons of continuous variables, normality was assessed within each group using the Shapiro-Wilk test. When normality was supported in both groups (p $>$ 0.05), group differences were evaluated using independent-samples t-tests. Homogeneity of variances was assessed using Levene’s test, and, when violated, Welch’s correction was applied. When at least one group deviated from normality, the Wilcoxon rank-sum test was used. Categorical variables were compared using Pearson’s chi-squared test.

To evaluate the association between reading-derived measures and cognitive impairment status, multivariable generalized linear models (GLMs) with a binomial distribution and a logit link function were fitted with CI vs HC status as the dependent variable. Separate models were estimated for each reading-derived predictor, including TRE, TCE, WPM, Total Reading Time, and the specific reading and comprehension error-type measures, while adjusting for gender, age, education, GDS score, Barthel Index, and CRIq. Model results are reported as odds ratios (ORs) with 95% confidence intervals (CIs). To account for multiple testing across these models, p-values were adjusted using the false discovery rate (FDR) procedure.

For each GLM, discriminative performance was further assessed by receiver operating characteristic (ROC) analysis, reporting the area under the curve (AUC) together with sensitivity, specificity, and accuracy at the selected probability threshold.

Finally, partial Spearman correlation analyses were conducted to explore relationships between reading-related measures and neuropsychological, functional, affective, and cognitive reserve variables, adjusting for age, gender, and education. A non-parametric approach was chosen because several variables were not normally distributed. The correlation analysis included audio-based and eye-tracking-based reading measures; eye-tracking variables were available only in a smaller subsample (n = 50). To account for multiple testing, p-values were adjusted using the FDR procedure, and only correlations that remained significant after FDR correction (q 0.05) were considered meaningful.

Exploratory eye-tracking analyses were also performed on the subsample with usable oculometric data following calibration and validation. In addition to eye-tracking timing measures, traditional first-order oculomotor metrics were examined. Statistical tests were selected based on assessments of normality and homogeneity of variance. Due to the smaller subsample size and the exploratory nature of these analyses, findings related to conventional eye-tracking measures were interpreted with caution.

All analyses were conducted in R (version 4.5.2; R Core Team, R Foundation for Statistical Computing, Vienna, Austria) [52]. Tables were generated using the gtsummary package (version 2.5.0) [53], and figures were produced using ggplot2 (version 4.0.2) [54]. Statistical significance was set at $\alpha$ = 0.05 (two-sided).