Can you detect early dementia from an email? A proof of principle study of daily computer use to detect cognitive and functional decline

Objective To determine whether multiple computer use behaviours can distinguish between cognitively healthy older adults and those in the early stages of cognitive decline, and to investigate whether these behaviours are associated with cognitive and functional ability. Methods Older adults with cognitive impairment (n = 20) and healthy controls (n = 24) completed assessments of cognitive and functional abilities and a series of semi‐directed computer tasks. Computer use behaviours were captured passively using bespoke software. Results The profile of computer use behaviours was significantly different in cognitively impaired compared with cognitively healthy control participants including more frequent pauses, slower typing, and a higher proportion of mouse clicks. These behaviours were significantly associated with performance on cognitive and functional assessments, in particular, those related to memory. Conclusion Unobtrusively capturing computer use behaviours offers the potential for early detection of neurodegeneration in non‐clinical settings, which could enable timely interventions to ultimately improve long‐term outcomes.

informant. Such tools are not ideally suited to detecting subtle changes in an individual's functional ability in everyday settings, over a prolonged period of time. 7,8 The challenge, therefore, is to detect objective and meaningful functional changes in higher-level IADL as early as possible and in ecologically meaningful settings, such as in the person's own home.
Capturing information about daily personal computer use activities may provide an opportunity to assess subtle changes in functional ability in elderly people over time. While personal computer use is an IADL in its own right, it also enables the user to complete a range of other complex IADLs, such as shopping, managing finances, and communicating. 9 The number of adults aged over 65 years using technology in the UK is increasing. Daily computer use in this age-group rose from 9% in 2006 to 45% in 2015, 10 accessing the internet on a mobile phone grew from 3% in 2011 to 21% in 2016, 11 and shopping online increased from 16% in 2008 to 45% in 2016. 11 Furthermore, as competent computer use relies on intact cognitive functioning across several domains (eg, attention, working memory, and executive function), changes in patterns of computer use (ie, functional change) may be a particularly sensitive indicator of cognitive decline. 12 Previous studies have demonstrated the feasibility of measuring computer use behaviours in older adults to distinguish between those with and without cognitive impairment. For example, it has been shown that people with MCI have reduced frequency and duration of daily computer use, 13 and take longer to complete an online questionnaire. 14 Seelye and colleagues 7 have also demonstrated that people with MCI make significantly fewer mouse movements, take longer pauses between movements, and have a higher variability in the trajectory of mouse movements. These behaviours were significantly correlated with cognitive test scores. Vizer and Sears 15 also demonstrated that keystroke speed and linguistic content is associated with cognitive impairment in older adults. In spite of these promising findings, it remains uncertain whether these individual computer use behaviours (eg, speed of use, typing abilities, and mouse operations) could be used as a composite marker of cognitive impairment in a single participant group. This is particularly important because a range of different behaviours are required to correctly operate a computer, and any one of these could be affected by cognitive decline. Another uncertainty in the field arises from the inclusion of novice or non-computer users in the participant sample of previous studies (eg, Kaye et al 13 ), which may limit the interpretation of findings due to the additional cognitive burden of learning to use a computer for the purposes of the study. Finally, the relationship between functional ability reflected by personal computer use and paper-based IADL measures has yet to be explored.
The study presented here is a cross-sectional proof of principle study designed to determine (1) whether multiple computer use behaviours, displayed by a sample of experienced older computer users on commonly undertaken computer tasks, can be used to distinguish between cognitively healthy older adults and those in the early stages of cognitive decline; and (2) whether these computer use behaviours are associated with cognitive and functional ability.

| Participants
Twenty participants with cognitive impairment (MCI, n = 17; mild dementia due to AD, n = 3) were recruited through the UK dementia research registry "Join Dementia Research", as well as through local memory clinics and community groups. Participants referred from memory clinics had all received a clinical diagnosis from a qualified memory specialist based on Peterson's criteria 16 for MCI or NINCDS-ADRDA criteria 17 for AD. Participants who self-referred to the study all reported a diagnosis of MCI or mild dementia due to AD, given by a specialist memory clinic. Specific clinical subtypes of MCI (ie, amnestic vs non-amnestic; single vs multiple domain) could not be ascertained. All participants had high functional ability, according to Katz criteria (all ≥5). 18 Twenty-four healthy control participants who had no prior history of cognitive impairment also participated in the study and were recruited through Join Dementia Research and local community groups (see Table 1

| Procedure
Participants were invited to take part in a single testing session lasting approximately 2 hours conducted either in their own homes or at The University of Manchester.

Key points
• This is one of the first investigations to explore a link between combined computer use behaviours and paper-based instrumental activities of daily living.
• A profile of computer-use behaviours can be used to differentiate between older adults with cognitive impairment and cognitively healthy older adults.
• Unobtrusively capturing data about various personal computer use behaviours could in the future be used to detect subtle, yet significant changes in cognitive and functional abilities.

| Cognitive and functional measures
Descriptive measures of global cognitive status were obtained using the Addenbrooke's Cognitive Examination (ACE)-III. 19 This test assesses 5 cognitive subdomains: attention, memory, verbal fluency, language, and visuospatial abilities, which provide a cognitive score out of a maximum of 100. Given that the only performance-based measure of executive function on the ACE-III is verbal fluency, we also incuded Part B of the Trail Making Test in the test battery as a measure of visual attention and task switching abilities. 20 Subjective ratings of cognitive and functional capacity were obtained using the Everyday Cognition (ECog) scale. 21 This assessment requires participants to rate their current functional abilities compared with 10 years previously. The 39-item questionnaire assesses cognitively based functional items, across 6 domains: memory, language, visuospatial abilities, planning (executive functioning), organisation (executive functioning), and divided attention (executive functioning). Scores range from 1 ("Better or no change") to 4 ("Consistently much worse"). To ensure high accuracy and detail of ECog ratings for cognitively impaired individuals, this test was com- Each group's mean total ACE-III and ECog scores and mean scores for each cognitive domain (including Trail Making Test Part B) can be seen in Table 1.

| Tasks of computer performance
All tasks assessing computer use performance were completed on a laptop (Lenovo Think Pad T540P) running Windows 7, 8, or 10, depending on which operating system the participant was familiar with from their own personal computer. Participants were provided with a separate keyboard and mouse if they preferred.
Participants were asked to follow a set of written instructions in order to complete 4 experimental computer tasks: (1) a basic Desktop navigation task, which included using the date and time function, use of folders, and the recycle bin; (2) a Word processing task that involved editing a Word document and writing a diary entry; (3) an email (Outlook) task that included opening, writing, sending, and deleting emails; and (4) an internet browsing (Internet Explorer) task that included performing a Google search and navigation of a webpage.
Participants could follow the instructions verbatim or adopt their own methods to complete the tasks, if they preferred.
Participants initially completed a practice session that involved shorter versions of the experimental computer tasks. The practice activity was repeated until the participant was confident in completing the tasks (approximately 2 repeats).  to higher-level linguistic and semantic features from more general operations, we analysed these separately and termed these "Text"

| Computer use behaviour data capture
and "Operational" keystrokes, respectively. Mouse operations included information such as total mouse clicks and the time, distance, and screen areas crossed.

| Statistical analysis
Outliers for each computer use variable were removed using the nonrecursive procedure 24 for each group of participants. This equated to 3.5% and 4.5% of data removed for the cognitively healthy control and the cognitively impaired groups, respectively. The distribution of the data was assessed using skewness and kurtosis. For non-normally distributed variables, the data were log transformed.

| Overall performance time variables
Compared with participants in the control group, cognitively impaired participants took longer to complete the computer tasks, paused more frequently overall and per minute, and had a longer total pause length per minute. By contrast, the mean duration for each pause did not differ significantly between the 2 groups. Therefore, the number of pauses per minute was chosen as the focus of further analysis based on the assumption that the greater total pause length per minute for the cognitively impaired group is due to them taking more pauses (of similar duration to control participants) per minute.

| Keyboard-use variables
Cognitively impaired participants made fewer "Text" keystrokes in total and per minute than the cognitively healthy participants. Because all participants took approximately the same length of time to complete the task involving "Text" keystrokes (approximately 3 minutes per participant), so total Text keystrokes and Text keystrokes per minute are a similar measure. Therefore, we focussed our analysis on Text Keystrokes per minute (ie, speed of typing). The cognitively impaired group did not differ significantly from the control group on total "Operational" keystrokes, but produced significantly fewer "Operational" keystrokes per minute. This difference was due to the different speeds the participants took to complete the tasks overall (see Section 3.1.1), and thus no further analysis was conducted on "Operational" keystrokes.

| Mouse-based variables
The cognitively impaired group executed a significantly greater number of mouse clicks compared with the control group, but there were no group differences on the number of clicks per minute. We selected total mouse clicks for further analysis based on the assumption that this indicated cognitively impaired older adults made more mistakes and then had to perform more clicks to correct these errors and therefore also contributing to the longer total duration to complete the tasks (see Section 3.1.1). The time between clicks (ie, inter-click interval) did not differ between the 2 groups. Mouse movements did not differ between the groups, as ascertained by the total number of pixels (ie, screen area covered) and the screen pixels within inter-click intervals (ie, speed of mouse movements).

| Correlations between computer use variables
Separate Kendall's Tau correlation analyses were conducted between the computer use variables selected from the group comparisons and each of the cognitive (ACE-III and Trail Making Test Part B; Table 3) and functional (ECog;   correlations were found (all P < .05), but only the Memory domain of the ACE-III and the ECog tests were significantly correlated with all 3 of the computer use variables.
Given that only the Memory domains were significantly correlated with all 3 computer use behaviours, we only included this cognitive domain within the regression models (Table 5). For mouse clicks and pauses per minute, neither age nor computer use experience could account for performance on these measures (all P > .05); however, the addition of ACE-III and ECog Memory scores led to a significant increase in the explained variance (R 2 change values both P < .05), and this model showed significant predictions of number of pauses per minute and number of mouse clicks (both P < .05). For "Text" keystrokes per minute, computer use experience was a significant predictor of performance accounting for 36.8% of the variability, which increased significantly with the addition of age (R 2 change = .101, P = .011) and increased significantly again with the addition of ACE-III and ECog Memory scores (R 2 change = .103, P = .020). Therefore, ACE-III and ECog Memory scores are significant predictors of keyboard typing speed (R 2 = .260, P = .003), but age and computer use experience may also account for variability in this behaviour.

| Accounting for within-group differences
To account for the possibility that the between-group differences were driven by those with mild dementia due to AD, all statistical analyses were repeated comparing only MCI participants to control participants. The results were unaffected, with the exception of ACE-III Memory score, which was no longer significantly related to number of mouse clicks.

| Receiver operating characteristic curve (ROC) analysis
The ROC analyses (Table 6) Table 6. When all the selected computer use variables were combined into a single predictive probability and compared with combined ACE-III Memory score, ECog Memory score, and Trail Making Test B predictive probability, correct classification was significantly higher for the combined computer use variables (z = 2.002, P = .045).

| DISCUSSION
In this proof of principle study, we examined whether computer use behaviours recorded from semi-structured tasks could discriminate

| CONCLUSION
This proof of principle study has demonstrated that a computer-based monitoring system can differentiate between cognitive impairment (ie, MCI and early AD) and healthy cognitive ageing using semi-directed computer tasks and several objective measures of computer use performance. The next phase will be to determine whether we can passively detect early changes over time in these same computer use behaviours, using unobtrusive recording of the behaviours through software embedded in participants' personal computers.
The ultimate aim is to ascertain whether behaviour changes associated with cognitive and functional decline could provide a sensitive and efficient way to detect very early signs of dementia.