We studied 99 elderly followed up at a Psychogeriatric outpatient clinic linked to a university hospital, aged between 60 and 90 years, who presented cognitive complaints or previous diagnosis of cognitive deficits, and a group of elderly engaged as volunteers in studies concerning cognition.

Individuals enrolled in the study were classified into three groups, paired according to age and educational level:

First, all participants were evaluated by the Montreal Cognitive Assessment (MoCA) (Nasreddine et al., 2005; Memória et al., 2013) for general cognitive screening, the Geriatric Depression Scale-15 (GDS-15) (Sheikh and Yesavage, 1986; Almeida and Almeida, 1999) to detect depressive symptoms, and the Lawton and Brody Scale for Daily Life Activities (L and B) (Lawton and Brody, 1969; Santos and Júnior, 2012) to assess functional independence for instrumental activities of daily living. Subsequently, the participants were submitted to a neuropsychological assessment, composed of the following tests: Rivermead Behavioral Memory Test (RBMT) (Wilson et al., 1989; Yassuda et al., 2010); Trail Making Test (TMT) (Spreen and Strauss, 1998; Campanholo et al., 2014); Digit Span (DS) (Wechsler, 1997; Nascimento, 2004), FAS-COWA (Spreen and Strauss, 1998); Rey Auditory Verbal Learning Test (RAVLT) (Rey, 1964; Malloy-Diniz et al., 2007), and Rey-Osterrieth Complex Figure (ROCF) (Osterrieth, 1944; Oliveira et al., 2004).

MCI patients were selected based on the criteria of Albert et al. (2011), which cover: (1) cognitive complaint, preferably confirmed by a relative or close person; (2) objective cognitive impairment in one or more cognitive domains, with performance below the expected for peers in the same age range and educational level; (3) not being demented, with preserved simple activities of daily living.

This group was further subdivided into aMCI (evidence of episodic memory impairment) and naMCI (evidence of impairment in cognitive functions other than episodic memory). The inclusion criteria for the MCI subgroups were performance at least 1.5 SD below the mean score on one function and/or between 1 and 1.5 SD below the mean score in two neuropsychological tests of the same function (Petersen, 2004).

All patients diagnosed as MCI had their blood counts, biochemical and lipid profile, vitamin B12, folic acid, syphilis serology, and thyroid function tested and performed magnetic resonance imaging studies to rule out metabolic, infectious, and vascular etiologies for cognitive decline.

Participants with no previous complaints of cognitive decline and who performed normally on cognitive tests were allocated to the GC.

An experienced multidisciplinary team that included neurologists, geriatricians, neuropsychologists, and speech-language pathologists gave the final diagnosis for the groups.

The exclusion criteria for all groups were: being illiterate; having a health condition that could preclude the realization of neuropsychological tests (such as, non-correctable vision or hearing impairment); history or evidence of cerebrovascular injury, non-AD dementias, and other neurologic/psychiatric conditions that might impair cognition (such as, severe traumatic brain injury, epilepsy, depression, bipolar disorder, and psychosis).

The local Ethics Committee approved the study under protocol number CEP 3.318.162, and all participants or their proxies signed a consent form before enrollment in the study.

After cognitive evaluation, the individuals were submitted to a protocol designed to assess inferential, semantic, and verbal episodic memory abilities.

Textual inference comprehension was assessed using The Implicit Management Test (IMT) (in the original French version, La Gestion de l'Implicite) (Duchêne May-Carle, 2000), adapted to Brazilian Portuguese (Silagi et al., 2014). The test evaluates the ability to comprehend inferences during reading activities and is designed to assess adult subjects with cognitive and/or communication complaints. The test was applied in a reduced-version containing ten texts consisting of short stories involving two people or describing a verbal interaction; individuals must read and answer three questions for each text. The only admissible answers are: “Yes”, “No”, or “I cannot answer.” The texts contain explicit and implicit information, which is necessary for the correct interpretation during reading. The texts were available for the patients to consult while they answered the questions.

Questions regarding the texts are subdivided into five categories that require different types of inferential reasoning: (1) explicit–questions are answered by using information supplied in the text; no inference-making needed; (2) logical–questions are answered by using a cause-effect relationship with the information provided in the text, through formal reasoning; (3) distractor–questions that have “I cannot answer” as the only possible correct answer, as the information required for an appropriate answer cannot be extracted from the text either explicitly or implicitly; (4) pragmatic–questions are answered by using context and previous experience; and (5) “other”–questions require both logical and pragmatic reasoning. Examples of the different types of questions can be seen in Table 1.

Semantic tests included: (a) the Word-Picture Matching (WPM) Test (Weintraub, 2000). This test evaluates spoken word recognition and assesses the frequency of semantic errors in word comprehension. The stimuli consist of five displays, each one containing pictures of four objects that are semantically related. Each display is presented four times (once for each picture as the target) in a total of 20 trials. The presentation order of displays is pseudo-randomized so that no four-picture display appears in sequential trials. One point is given for each correct response given at the first attempt, which allows a maximum score of 20—“I do not know” is considered as an incorrect response; (b) the Semantic Associates Test (SA) (Weintraub, 2000). This test comprises16 displays containing three items presented as two pairs (one target and one distractor), which evaluates the functional, contextual and categorical relationship between the items. The individual is asked to look at the two pairs of pictures in each display (semantically related or non-related) and point to the matching set. For example, in a display containing the pairs: sweater-blanket/sweater-pillow, the correct answer is sweater-blanket (functional relationship); for the two pairs sweater-chest/sweater-workbench, the correct answer is sweater-chest (contextual relationship). One point is given for each correct match choice, which allows a maximum score of 16.

To assess verbal and non-verbal episodic memory, we used the Three-Words Three Shapes Test (3W3S) (Weintraub, 2000). The participants are asked to copy three words (pride, hunger, station) and three geometric shapes displayed on a sheet of paper and then reproduce them from memory to assess immediate and delayed recall (after five minutes, during which the individuals performed a task of picture description). The test provides scores for verbal memory (immediate and delayed recall for words; WImR and WDR, respectively) and non-verbal memory (immediate and delayed recall for shapes). We considered only the verbal memory scores in this study: each word correctly retrieved in each condition is scored as five, giving a total score of 15. The score of five for each correct word allows for deductions due to adding, subtracting, or substitution errors in writing.

Intergroup comparison of means for demographic and clinical continuous variables was performed through one-way analysis of variance with Bonferroni's post-test. The Chi-Squared test was used to assess intergroup differences in sex distribution. Mixed-effects linear regression models were used to (a) determine the best predictors for IMT total scores and scores in each question subtype in the total sample, using such scores as dependent variables, schooling and scores on WPM, SA, subtests of the 3W3S as fixed-factors, and within-subjects differences in performance as a random effect; (b) determine the best predictor for IMT in each diagnostic subgroup using diagnosis as the dependent variable, schooling, and scores on WPM, SA, subtests of the 3W3S as fixed-factors, and within-subjects differences in performance as a random effect. Schooling was tested in interaction with all variables. The Akaike information criterion (AIC) was employed to compare the models and choose the best among them, which were those with the smaller AIC values. Model estimations were performed using restricted maximum likelihood (REML). Significance levels were set at p < 0.05 values. We used the SPSS^® Statistics software version 25 for all analyses.

Regarding the performance of the groups in the inference comprehension test (IMT), we verified that the MCI group showed worse performance than controls in the total score. The aMCI group performed poorer than controls in “logical” questions but with a p-value approaching the limit of statistic non-significance. The aMCI and naMCI groups performed similarly regarding total scores and subtypes of questions.

Few studies in the literature have addressed the performance of individuals with MCI in inference comprehension tasks. We found three studies that showed similar results about the difficulty of understanding inferences in subjects with MCI. Schmitter-Edgecombe and Creamer (2010) verified that aMCI subjects produced fewer inferences in a story comprehension task than controls and had more difficulties explaining story events and using preliminary text information to support inference generation. Similarly, Gaudreau et al. (2015) found that MCI participants were impaired in identifying ironic or sincere stories that required mental inference capacities, compared to control subjects. We found only one study that compared aMCI and naMCI on the ability to understand inferences. Silagi et al. (2021) evaluated a different cohort of MCI individuals with the same test used in our study (IMT). They found that MCI patients had difficulty understanding inferences compared to controls and the accuracy analyses showed that the total score in the IMT provided good sensitivity and specificity in discriminating MCI from normal individuals. However, they were also unable to differentiate the MCI subgroups from each other in the task.

As for the semantics tasks, all groups showed similar performance in the WPM and SA tests. We observed a ceiling-effect regarding performance in the WPM task in all groups. Semantic impairment has been widely reported in MCI in tasks ranging from verbal fluency, naming, and sentence comprehension (Emery, 2000; Balthazar et al., 2008; Rinehardt et al., 2014; Silagi et al., 2015). However, studies show that the performance on isolated word comprehension and semantic association (as required in WPM and SA) are usually preserved in the patients until the early stages of AD, that is, the semantic impairments in MCI are associated with difficulty in lexical or lexico-phonological search, and in complex tasks involving a more fine-grained semantic decision, with basic semantic knowledge preserved (Ortiz and Bertolucci, 2005; Barbeau et al., 2012; Kirchberg et al., 2012; Tsantali et al., 2013; Venneri et al., 2018).

On the other hand, the verbal episodic memory tests (3W3S) were able to differentiate between controls and MCI as well established in the literature (Ding et al., 2019; Wasserman et al., 2019; Abraham et al., 2020; Silva et al., 2020).

Our mixed-effect model showed that schooling and verbal episodic memory tasks were the best predictors for inference-making performance in the whole sample. Regarding the predictors of the ability to understand inferences in different diagnostic groups, we found distinct profiles in healthy controls and MCI groups.

In the control group, schooling and WPM were the best predictors, with a trend for SA. A study with healthy subjects showed a strong effect of education on inference comprehension; individuals with higher educational levels had better performance than individuals with a lower educational level on the total score and across all question types of the inference test (Silagi et al., 2014). The authors discussed that schooling plays an essential role in developing several cognitive-linguistic skills, which support inferential reasoning (Ardila, 2005).

The influence of semantic knowledge was also proven in the ability to understand texts with inferences in healthy subjects (Yeari and van den Broek, 2015, 2016; Dong et al., 2020). Semantic integration is essential for inference making, and it occurs as a result of the activation of the semantic network that produces access to a word's meaning with consequent co-activation of words and meanings of the same semantic category. Inference-making is based on the existence of strong semantic associations with specific ideas conveyed by speech or text. Yeari and van den Broek (2015) found that robust associations between textual information and personal background knowledge result in a greater probability of the information being activated. This activation of inference from background knowledge is a function of the number and strength of semantic constraints imposed by the evoking text.

In the aMCI group, the best predictor of inference comprehension was verbal memory. The influence of episodic memory on the ability to understand inferences is well established and addressed by previous studies in subjects with MCI. Schmitter-Edgecombe and Creamer (2010) reported impairment in the production of inferences in MCI patients associated with poorer delayed verbal memory abilities. The authors argue that episodic memory influences inference comprehension because it helps create causal connections between different text parts and establish an integrated story. Similarly, Gaudreau et al. (2015) showed that the poorer comprehension of mental state inferences in the MCI population was correlated with episodic memory and executive functions difficulties. Finally, Silagi et al. (2021) found correlations between the scores in the inference comprehension task and RAVLT in an aMCI group, presupposing that the impairment in inference comprehension observed in this group was related to episodic memory failure. In contrast, they found correlations between the total test score and the TMT-A in the naMCI group, associating the poor inference comprehension in this group with failure in attention and executive functions (which we did not include in our mixed-effect models).

In our results, in the naMCI group, inference-making skills were associated with the semantic tasks (WPM and SA). According to Graesser et al. (1994), some types of inferences are built when long-term memory knowledge structures are activated and incorporated into the representation of the text's meaning. Van Dijk and Kintsch, 1983 establishes the relationship between inferential processing and how information is stored, highlighting that long-term memory is organized in semantic bundles, hierarchically ordered. Lastly, Rumelhart (1980) defines schema as the data structure for representing concepts stored in memory. Schemas represent knowledge about concepts, events, and actions. According to the author, the schema underlies a concept stored in memory, which generates meaning. Schemas allow access to information stored in memory while reading and building inferences.

As for the different types of IMT questions, we hypothesize that the main predictors reflect the nature of reasoning necessary to successfully identify and extract the information in order to answer the questions related to the texts:

This study belongs to a series in which we explore high-demanding language abilities, including inferential processing (both textual and visual) in the normal aging-AD continuum and stroke patients (Silagi et al., 2014, 2018, 2021). Based on previous results in normal individuals and MCI patients, we traced some variables influencing performance in the IMT, including executive functions and episodic memory (Silagi et al., 2014, 2021). We then decided to proceed with our investigation focusing on the role of semantic abilities in inference-making by employing the WPM and SA tasks. An interesting finding in our study was that by conducting a complementary analysis of the “justifications” (that is, how the stimulus word “matched” the individual's response) in the SA task, we observed that many patients could not explain it or provided nonsense explanations, despite making the correct choice (data not shown). We believe that these observations point to a subtle impairment of semantic processing, albeit not sufficient to prevent a correct choice. We intend to explore these findings in subsequent studies.

This study has some limitations: its cross-sectional design and the need for a larger sample to include a higher number of variables in our predictive model. The latter was precluded by the Covid pandemic that impeded us from proceeding with the enrollment of patients. Another limitation is that, given the ceiling-effect observed in WPM, we could have obtained better results regarding the impact of this semantic ability on inference-making should we have used a more demanding task. More studies, preferably longitudinal, are needed to confirm the results.