A number of studies examined differences between speech and writing associated with aging (Croisile et al., 1996; Mitzner and Kemper, 2003; Forbes-McKay et al., 2013, 2014). Croisile et al. (1996) compared oral and written descriptions of a picture commonly used in the assessment of aphasia [the Cookie Theft picture (Goodglass et al., 2001)]. Of the 22 patients with probable AD and their 24 matched controls, patients produced fewer words across all word categories, fewer predefined information units (also called pictorial themes), and sentences that were grammatically less complex (e.g., fewer subordinate clauses). For either group, no significant difference between oral and written modalities was found in the use of pictorial themes. However, the increase of irrelevant information and the decrease in syntactic complexity in the written task led the researchers to conclude that the written mode was more sensitive for identifying linguistic difficulties present in AD.

Forbes-McKay et al. (2013, 2014) also conducted a cross-sectional and longitudinal study on speech and writing to document the progression of linguistic impairments in AD over a 12-month period. Thirty-one probable AD patients and 30 matched controls completed a simple and a complex narrative description task orally and through handwriting. Cross-sectional comparisons indicated that the language of mild-moderate AD patients was characterized by a higher number of empty and indefinite phrases, fewer pictorial themes and error corrections, and shorter and grammatically more simplified sentences in both the oral and the written mode. Follow-up data suggested that both phonological and visual processing declined over time. In these studies the data were analyzed manually.

Toledo et al. (2014) added an automated component to these analyses. They used a simple (“The Tripping Woman Picture”) and a complex line drawing (“The Traffic Chaos Picture”) that were also used by Forbes-McKay et al. (2013, 2014). Their first goal was to develop machine learning classifiers to classify healthy subjects based on years of education. For this, linguistic features, such as words per sentence and incidence of nouns, verbs and adjectives, were extracted from handwritten descriptions with the use of natural language processing tools. Their second goal was to automatically identify the features that were most suited to distinguish the groups. Results showed that, for data classification, a specific type of machine learning (i.e., Support Vector Machine with a radial basis function kernel) performed best. Moreover, correlation-based feature selection was the preferred method to replace manual selection with (Toledo et al., 2014). These findings also support the idea that groups can be distinguished on the basis of specific linguistic properties.

Not only written texts (product), but also keyboard interactions (process) can indicate cognitive (and motoric) status. For example, machine learning research has shown that keyboard typing can detect cognitive stress (Vizer et al., 2009; Unni et al., 2021). In a study by Vizer et al. (2009) keystroke features (e.g., time per keystroke) and linguistic features (e.g., lexical diversity) of free text production were identified. When used to classify a cognitive stress condition relative to a non-stress condition the classification accuracy amounted to 75% (i.e., a percentage also obtained in affective computing methods). This indicates that keyboard interactions might be able to track changes in cognitive stress related to both aging and AD.

This is supported by copy typing studies, such as a study by Van Waes et al. (2021) that showed a non-linear relationship between typing speed and age. Participants from 13 to 83 years old performed a copy task in which they had to copy type letters, sentences and words from a screen. The non-linearity was present for all subtasks, but differed for lexical (e.g., typing of sentences) compared to non-lexical subtasks (e.g., tapping task). Overall, participants typed fastest between the ages of 21 and 30, but after that their speed started to decrease. In addition, subtasks with lexical components showed a gradual increase in typing speed between the ages of 13 and about 25, while this was not the case for non-lexical subtasks. In the older age groups (from about 50 years old), typing speed also decreased more in lexical than in non-lexical subtasks (Van Waes et al., 2021). Another study with the same copy task showed that keyboard typing can also detect gradual changes in motor performance related to AD (Van Waes et al., 2017). Young adults, healthy elderly and age-matched impaired elderly performed the copy task. Results showed not only a decrease in typing speed with increasing age (i.e., between young adults and healthy elderly), but also a slowdown for cognitively impaired elderly compared to healthy elderly. These observations suggest that aging affects the skills required for lexical processing and further support the idea that a typing task could be promising for screening and follow-up of Alzheimer patients − especially because the type of task allows to measure a combination of motor and cognitive capacities in language production (Van Waes et al., 2017, 2021).

In writing and in speech, pausing behavior has proven to be an important resource for identifying cognitive effort during text composition (Wengelin, 2006; Vizer et al., 2009). Pausing behavior points to production flow interruptions [the term often used here is “fluency” (Hayes and Chenoweth, 2007; Van Waes and Leijten, 2015)]. This can be measured, for instance, in terms of the number and length of pauses. According to capacity theory, interruptions result from exceeding the limited capacity of working memory (Just and Carpenter, 1992; McCutchen, 1996). The large number of concurrent writing processes that take place during writing cause a cognitive overload. Subsequently, the writer pauses so that working memory resources are freed up and can be used for other, more demanding processes (e.g., planning, formulating or reviewing). Once enough processes are completed, writing can resume (Kellogg, 2001; Olive and Kellogg, 2002; Olive, 2014).

In tandem with pauses, also the number and length of production bursts change with cognitive effort. When the number of parallel processes reduces, the writing process shows more fragmentation (i.e., more switching between pausing and writing), characterized by longer pauses and shorter bursts. When the number of parallel processes increases, this pattern changes to less fragmentation with shorter pauses and longer bursts. This was demonstrated by studies examining the influence of transcription automation on the distribution of pauses and bursts during writing. The cognitive load of non-automatised transcription can, in fact, be quite high. Studies in children and adults showed that writing processes are less fluent when typing is less automated, which is reflected by more and longer pauses and shorter bursts (Alves et al., 2007; Alves and Limpo, 2015; Leijten et al., 2019).

In healthy aging, these disfluencies are especially evident in the form of longer pause times caused by word retrieval failures such as tip-of-the-tongue states (Burke and Shafto, 2004; Verhaegen and Poncelet, 2013). This decline in working memory capacity with age was also suggested by studies using a copy task; a task that minimizes the load of cognitive processes such as planning, formulating and reviewing on working memory. A study by Van Waes et al. (2021) showed higher interkey latencies for older participants. In addition, the influence of age was more pronounced in lexical than in non-lexical tasks, an effect that was likely caused by the added load of formulating on working memory (Kliegel and Jäger, 2006; Wild-Wall et al., 2011; Van Waes et al., 2021). With regards to AD, patients tend to be less fluent writers as well, reflected by shorter production bursts and more pauses. Moreover, they pause longer before and within words than their healthy counterparts (Leijten et al., 2015).

Research into the development of language-based screening instruments for AD is scarce and rather limited in scope. First, it almost exclusively examines the characteristics and progression of cognitive decline on the basis of product measures (i.e., based on the final texts; McGeown et al., 2009; Forbes-McKay et al., 2013, 2014; Toledo et al., 2014). This is also reflected in most studies focusing on writing in elderly, such as those focused on text analysis of healthy elderly (e.g., Kemper, 1990; Ferguson et al., 2014) or of people with AD (e.g., Snowdon et al., 1996; Garrard et al., 2005; Pakhomov et al., 2011; Forbes-McKay et al., 2014). In this paper, we want to add to these studies by exploring the possibilities of writing process dynamics to identify cognitive decline.

Second, language-related assessment tools in the diagnostic work-up of AD have hardly changed over the last 30 years. However, the information technology landscape has changed considerably. Especially the increase of digitalization in written text production is remarkable (Brandt, 2014), and the group of elderly with computer literacy is continuously growing (Jones and Fox, 2009; Zhou, 2011). This paper will therefore look at typed rather than handwritten texts. An instrument that can detect cognitive decline based on digital writing processes (i.e., captured with keystroke logging) would be easy to apply, cheap, unobtrusive, non-invasive, and adapted to the current and future technological environment.

Since findings from previous research indicate the need for a screening task that assesses multiple language components (Forbes-McKay et al., 2014), we will test digital writing tasks that focus on motoric and cognitive aspects. To compare the cognitive aspects of text production between the participant groups, the following research questions will be addressed (using the variables listed under each question):

The participants were 30 subjects, distributed over two groups:

A Mann–Whitney U test showed no significant difference for age between both groups (U = 112.500, Z = 0, p = 1, r = 0). Moreover, all participants underwent a full neuropsychological examination at the time of the study, amongst others consisting of a Mini-Mental State Examination (MMSE; maximum score of 30 with a higher score indicating better cognitive functioning) (Folstein et al., 1975, 1983) and the Geriatric Depression Scale (GDS; maximum score of 30 with a high score indicating significant depressive symptoms) (Sheikh and Yesavage, 1986). The median scores on the MMSE and GDS were, respectively, 29 and 3 for the healthy controls and 26 and 8 for the cognitively impaired patients. The MMSE scores showed, in line with our expectations, a significant difference between both groups (mean ranks of healthy controls and cognitively impaired patients were 19.433 and 11.567, respectively; U = 171.500, Z = 2.948, p = 0.003, r = 0.548), as did the GDS scores (mean ranks of healthy controls and cognitively impaired patients were 10.233 and 20.767, respectively; U = 33.500, Z = −2.776, p = 0.006, r = −0.534).

During the experiment, a general questionnaire was completed to gather demographic information and to determine work experience, handedness, computer and typing skills. In addition, the examiner administered the MMSE and the GDS (see section “Participants”) to determine participants’ cognitive status. Finally, participants completed the main tasks consisting of two separate picture description tasks.

All participants, when seated at a computer, were instructed to write a short descriptive text for each of two picture description tasks: the Cookie Theft picture from the Boston Diagnostic Aphasia Examination (Goodglass et al., 2001) and the situational drawing from the Dutch version of the Comprehensive Aphasia Test (Swinburn et al., 2005; CAT-NL; Visch-Brink et al., 2014). To avoid order effects, the picture description tasks were counter-balanced. The pictures were included in a MS Word document (see Figure 2). The instruction (translated from Dutch) was as follows: “Describe what you see on this picture and what you think will happen.”

To observe writing processes on a computer, keystroke logging programs have been developed (Van Waes et al., 2012). These programs collect keyboard and mouse activities during text production. Because they also add a time stamp to each logged event, the resulting log files can be used to reconstruct and describe text production processes. In this study, data were collected with the keystroke logging program Inputlog (Leijten and Van Waes, 2013).

To facilitate the exploration of the cognitive activities during the writing process, the main focus in this study was on pausing behavior. Pauses reflect processes of activating conceptual and linguistic knowledge in order to accomplish specific rhetoric goals (Leijten et al., 2019). Inputlog offers some predefined analyses by default, with the so-called general analysis being the foundation. Apart from some metadata, the general analysis represents the logging process as a fine-grained dataset containing all keystrokes, mouse movements and their respective time stamps in a linear order. Each row in the output represents a keystroke or a mouse action, documented with position and time-based information [key-in and key-out time in milliseconds (ms)] that is used to calculate the pause time between two typing events (e.g., time between two consecutive keystrokes, measured from key-in to key-in). In the context of this study, three other analyses were used: the summary, pause and linguistic analyses (Leijten and Van Waes, 2020). These analyses represent the data of the general analysis on an aggregated level.

As described, the data output from keystroke loggers is mainly based on keystrokes and mouse movements, each captured as isolated events. Moreover, Inputlog aggregates the logged process data from the letter level (keystroke) to the word level as well (Leijten et al., 2015, 2019). Specifically for the linguistic analysis, this allowed us to merge and annotate the word-level output with existing lexica and to analyze these enriched process data with traditional Natural Language Processing (NLP) tools. In this procedure, the logged process data were thus supplemented with different types of linguistic information such as lemmata, part-of-speech tags, syllabification and chunking (see Leijten et al., 2015, 2019, for a more detailed explanation of the linguistic analysis of Inputlog).

First, we examined the effects of group on a number of overall process measures. Second, pause variables were observed at the task level, together with more specific process variables (i.e., pause times between words in general and for word categories) at the word level. Because the data were hierarchical, we used mixed-effects models (Baayen, 2008). For each variable we built models of increasing complexity, with the next model always building on the previous one. Gradually increasing the models’ complexities allowed us to test the impact of each additional factor (see Table 1).

For the overall process measures and the pause variables at the task level only the first two models were built. To test the influence of group on these variables, we first estimated the baseline and the group effect models. The baseline model included task and typing speed as fixed factors and participant as a random factor. In the group effect model, group was added as a fixed effect. To compare the models we used the likelihood ratio test and compared the AIC values of the models (Baayen, 2008).

For pause time between words, all four models were built. To test the influences of group and part-of-speech on pause times, we first estimated the baseline, the group effect and the part-of-speech effect models. The baseline model included task and typing speed as fixed factors and participant as a random factor. In the group effect model and the part-of-speech effect model we included group and part-of-speech, respectively. Finally, to test if the influence of group varies between different word categories a group-by-part-of-speech interaction was included in the interaction effect model. Here, we also used the likelihood ratio test to compare the AIC values of the models (Baayen, 2008).

Analyses were carried out in R with the packages lme4, car, and multcomp (Hothorn et al., 2008; Bates et al., 2015; Fox and Weisberg, 2019; R Core Team, 2019).

For time on task, results showed that the group effect model was the best fit when comparing the baseline and the group effect models [X² (Df = 1) = 5.591, p = 0.018; see Table 3]. Hence, there was a statistically significant main effect of group. In Table 7, the coefficient of group shows that patients needed 0.772 standard deviations (which equals 123.836 s) more time to describe the pictures than healthy controls. This means a medium to large effect was found (Sawilowsky, 2009). For the number of characters per minute, the group effect model was the best fit when comparing the baseline and the group effect models [X² (Df = 1) = 17.715, p < 0.001]. This means that also for this variable there was a statistically significant main effect of group. The coefficient of group shows that patients typed 0.566 units (i.e., 108 characters) less per minute (when pure typing speed is already taken into account) than healthy controls.

For number of pauses per minute, the group effect model was the best fit when comparing the baseline and the group effect models [X² (Df = 1) = 15.169, p < 0.001; see Table 4]. Hence, there was a statistically significant main effect of group. In Table 8, the coefficient of group shows that patients paused 0.490 units (i.e., 102 pauses) less often than healthy controls. For proportion of pause time, results showed that the group effect model was the best fit when comparing the baseline and the group effect models [X² (Df = 1) = 25.520, p < 0.001]. This means that also for this variable there was a statistically significant main effect of group. The coefficient of group shows that pause time in proportion to time on task was 1.356 standard deviations (i.e., 20.589%) higher for patients than for healthy controls, which means a very large to huge effect was found.

For number of P-bursts per minute, the group effect model was the best fit when comparing the baseline and the group effect models [X² (Df = 1) = 13.238, p < 0.001; see Table 5]. Hence, there was a statistically significant main effect of group. In Table 9, the coefficient of group shows that patients’ writing processes contained 1.710 more bursts than healthy controls’. For duration of P-bursts, results showed that the group effect model was the best fit when comparing the baseline and the group effect models [X² (Df = 1) = 26.269, p < 0.001]. This means that also for this variable there was a statistically significant main effect of group. The coefficient of group shows that patients’ P-bursts were 1.373 standard deviations (i.e., 4.926 s) shorter than those of healthy controls, which means a very large to huge effect was found.

Given the small sample, this study is only exploratory and should be confirmed in a larger cohort. Moreover, this study is performed cross-sectionally at a single moment in time. However, to develop a robust screening tool it should be repeated on several occasions during follow-up to document language change, as described in Forbes-McKay et al. (2013, 2014). We believe that, to date, no longitudinal assessment of the writing process characteristics of AD patients has been carried out to describe the progression of the disease.

Furthermore, there may have been an order effect in the keystroke data because the (neuro)psychological tests (i.e., MMSE and GDS) were administered before the picture description tasks. A study by Vizer et al. (2009) previously showed that keyboard interactions can detect cognitive stress. Hence, if the (neuro)psychological tests induced cognitive stress, it is possible that this was picked up by the typing tasks that followed. However, because we would expect it to affect both groups equally and because this is a cross-sectional study, we believe that it did not influence the results in this paper. For future research, however, it is recommended to start with the typing tasks instead of the (neuro)psychological tests, or to split the experiment into two separate sessions.

First, although the Cookie Theft picture by Goodglass et al. (2001) has been widely used in previous studies (e.g., Forbes-McKay et al., 2013, 2014), its task validity has, to our knowledge, not been reported. Moreover, the Cookie Theft picture dates back to 1983 and might not be fully adequate anymore in current times. During preliminary analyses, the situational drawing by Visch-Brink et al. (Swinburn et al., 2005; Visch-Brink et al., 2014) proved to have more discriminatory power than the Cookie Theft task (Goodglass et al., 2001). That does not pose a problem for this cross-sectional study because this was accounted for in the mixed-effects models. However, test materials in longitudinal studies should be congeneric and exchangeable (Jöreskog, 1971; Paesen and Leijten, 2019), at least with respect to the most distinguishing variables.

Second, in this paper we only focus on pauses related to eight specific word categories. However, pausing behavior is also dependent on the context in which these word categories occur and on specific features of words or other lexical units. For example, pauses before certain parts-of-speech have shown differences depending on the specific phrases in which they occurred (Leijten et al., 2019). Moreover, both word frequency and length have been shown to influence pause times, with more frequent and shorter words leading to shorter between word pauses before those words (Medimorec and Risko, 2017). Hence, in a follow-up study attention should be given to specific features of lexical units as well as the (linguistic) contexts in which these units occur.

Third, the present study only looks at texts in Dutch. However, specific language properties, such as basic word order [e.g., Subject-Object-Verb in Dutch versus Subject-Verb-Object in English (Bastiaanse and van Zonneveld, 2005)] or aspects of certain word categories [e.g., two definite articles (de/het) in Dutch versus only one in English (the)] may differ between languages. The results in this paper should, therefore, not simply be generalized to other languages and it is recommended that future studies explore the consistency of our findings in languages other than Dutch.

Fourth, the analyses in this paper did not account for changes in the writing processes over the course of the task. However, it could be that participants’ typing and/or pausing behavior changed over time because of factors such as fatigue or stress, due to a learning effect, or because they adapted their strategies to changes in the task situation (Rijlaarsdam and Van den Bergh, 1996; Vizer et al., 2009; Ulinskas et al., 2017; Xu, 2018). To take into account such fluctuations, future studies could treat the writing processes as time series in their analyses (Allen et al., 2016; Pham, 2018).

Fifth, since writing can be considered a combination of graphomotor, sensory and cognitive skills, in a follow-up study we strive to relate writing process data of a motor task (such as a controlled copy task), instead of a measure based on interkey latencies as was done for this study, to a more open writing task. Personal motor characteristics of each participant can then be treated as individual benchmarks for further analysis of the cognitive data. This is important, because interkey intervals differ considerably between participants (Grabowski et al., 2010; Wallot and Grabowski, 2013; Van Waes et al., 2019).

Finally, the process approach presented here should allow online written data to be described and analyzed from a longitudinal perspective. The relative simple applicability also enables writing process researchers to gather a large dataset of normative data from adults of a wide range of ages as to build a normative dataset (Ferguson et al., 2014; Saryazdi et al., 2018; Torrance et al., 2018). Since AD is likely to begin many years before the onset of symptoms (Snowdon et al., 1996; Taler and Phillips, 2008; Blazer and Steffens, 2009; Vermunt et al., 2019), an early, non-intrusive work-up creates added opportunities for screening of the disease and initiation of therapies (Blazer and Steffens, 2009, p. 249).