The participants were recruited from senior centers and daycare centers in Seoul and Asan, South Korea. The study included 2 groups: 64 healthy controls (HCs) and 47 patients with MCI. In accordance with a previous study [1], the inclusion criteria for MCI were as follows: (1) a subjective memory complaint; (2) memory impairment relative to age- and education-matched HCs, confirmed by performance on a neuropsychological battery (below 1.5 SD); (3) intact global cognitive function, confirmed by a score on the Korean version of the Mini-Mental Status Examination (MMSE-K); (4) independent activities of daily living; and (5) experience using a smartphone for at least 3 months. The exclusion criteria were as follows: (1) diagnosed with dementia by a clinician, (2) neurological or psychiatric disorders such as stroke and depression, and (3) visual or auditory impairments. These criteria are based on the original Petersen criteria, which restricts MCI to memory problems (amnestic MCI) only.

This study was approved by the Institutional Review Board of Soonchunhyang University (202306-SB-070-04) and registered at the Thai Clinical Trials Registry (TCTR20220415002). This study adhered to all relevant guidelines and regulations concerning the ethical conduct of human subject research. All participants provided written informed consent prior to the initiation of any study activities. The consent process ensured that all participants were fully informed about the nature of the study, its objectives, and their rights as participants, including the voluntary nature of their participation. To protect participants’ privacy and confidentiality, all study data were anonymized and deidentified before analysis. No identifying information was collected or stored that could be traced back to individual participants. The study followed strict data protection protocols to ensure the confidentiality of the information provided by the participants. All participants were compensated with ₩30,000 (~US $21.74) for their study participation. No images that could potentially identify individual participants were included in the paper or supplementary materials.

All participants performed the Korean version of the MoCA (MoCA-K) and the computerized Corsi block-tapping test (CBT) before their keystroke data were collected. Both the MoCA-K and the computerized CBT were used to assess baseline cognitive performance. Subsequently, the relationship between these assessments and keystroke dynamics was examined. The MoCA-K was selected for its role as a screening tool for MCI and its ability to assess global cognitive function, while the CBT was chosen for its efficacy in assessing working memory decline, which is common in patients with MCI.

Participants then installed a custom-developed mobile app that allows users to measure health conditions through typing on a smartphone. The app keyboard replaced the default keyboard, and participants were allowed to become accustomed to it for a week. During regular typing activities, keyboard interactions (keystroke dynamics) such as key presses and releases were recorded and time-stamped in the background. The keystroke dynamics were stored in a JSON format and indexed in a database, accessible only to the app. The app periodically delivered information on uniquely coded keystroke dynamics to a remote cloud server (Microsoft Azure), when the participant’s device was connected to Wi-Fi and charging.

Data collection lasted a month, and all participants were instructed not to have anyone other than themselves type on their phones. Although it was impossible to confirm whether participants typed all data definitively, the author reconfirmed that only the participants themselves used their phones for 1 month before data analysis. This study did not impose a minimum daily usage requirement on participants, as the aim of this study was to capture keystroke dynamics that most accurately reflect the natural status of their daily lives. Within a month, 3530 typing sessions were collected, 2250 (63.7%) from 64 HCs and 1280 (36.3%) from 35 patients with MCI. Of the 3530 typing session, 2740 (77.6%) with more than 40 key presses per session were finally used to ensure sufficient data for meaningful comparison, in accordance with a previous study [14]. The flow diagram of the overall procedure of this study is presented in Figure 1.

Two variables of the keystroke dynamics stored in the JSON files were analyzed in this study: hold time (HT) and flight time (FT). These 2 variables were chosen because a previous study showed that they are significant in discriminating MCI from healthy aging among other keystroke dynamics [13]. HT refers to the time interval between pressing and releasing a key, while FT indicates the time interval between releasing a key and pressing the next key (Figure 2). To preprocess data, HT >700 microseconds and FT >3 seconds were excluded according to the guidelines in a previous study [14].

The original MoCA was developed to discriminate MCI from normal aging, and its domains consist of visuospatial/executive function, attention, memory, language, abstraction, and orientation [20]. Its score ranges from 0 to 30 points, where higher scores indicate better global cognitive function. The MoCA-K was adapted based on the original MoCA [21] (Multimedia Appendix 1). The cutoff score of the MoCA-K for MCI was 23, and 1 point could be added for participants with fewer than 6 years of education. In a previous study with a cutoff score of 23, its sensitivity and specificity were 94.2% and 40.5%, respectively [21].

The computerized CBT was used to assess attention and working memory. Since patients with MCI commonly exhibit a decline in memory and executive functions [1], a working memory test that evaluates both memory and executive function components was chosen [22]. In the task, 9 white squares were randomly located on a tablet monitor. Some of the squares changed their color from white to red in a sequential manner while participants were asked to memorize the locations and sequences of the changed squares. Afterward, participants were instructed to point out the changed squares in the order in which their colors had changed by touching the screen [23,24]. In this study, the number of squares to be changed was 5, and 15 trials were implemented. A previous study indicated that the average square span of HCs is 5 [23].

The MMSE is a widely used tool for screening cognitive impairment. The MMSE-K was the result of the MMSE being translated into Korean and standardized [25] (Multimedia Appendix 2). It comprises 12 items assessing 7 cognitive domains: orientation, memory registration, attention and calculation, recall, language, understanding and judgment, and visual construction. Its score ranges from 0 to 30, with scores below 24 indicating cognitive impairment [25].

All outcome measures were implemented by an occupational therapist with 5 years of clinical experience.

SPSS for Windows (version 22.0; IBM Corp) was used to analyze the data. The demographic characteristics of the participants were analyzed using descriptive statistics. Independent 2-tailed t test and chi-square test were used to compare both groups. A receiver operating characteristic (ROC) curve analysis was used to confirm sensitivity and specificity, and a cutoff score for patients with MCI was determined according to the highest Youden Index (sensitivity + specificity – 1), which could be a criterion for choosing an optimal cutoff score. To conduct ROC curve analysis of the combination of keystroke dynamics as a single variable, predictive probability through logistic regression was used in accordance with a previous study [26]. A Spearman correlation was performed to examine the relationship between keystroke dynamics and global cognitive function and working memory. Statistical significance was set at P<.05.

There were no statistically significant differences in demographic characteristics between both groups (all P>.05) except for the score of the MoCA-K (P<.001). This finding indicated that HCs were age- and education-matched to patients with MCI except for cognitive function (Table 1).

For discriminating patients with MCI from the matched HC, both HT and FT showed a higher Youden Index than the MoCA-K (HT: .847; FT: .947; and MoCA-K: .469), suggesting that keystroke dynamics can better discriminate MCI compared to the conventional MCI screening tool (Table 2). Specifically, FT yielded a maximum sensitivity (97.9%) and specificity (94.7%). Interestingly, when combined with the MoCA-K, HT had a higher Youden Index, while FT did not (Table 2 and Figure 3).

Keystroke dynamics were found to be significantly correlated with the MoCA-K (HT: r=–.468; P<.001; FT: r=–.491; P<.001) and the CBT (HT: r=–.487; P<.001; FT: r=–.492; P<.001; Table 3). These findings suggested that keystroke dynamics are associated with global cognitive function and working memory.

The aim of this study was to evaluate keystroke dynamics as potential digital biomarkers for MCI. For this purpose, keystroke dynamics were investigated in 64 HCs and 47 patients with MCI, alongside the conventional screening tool. By measuring keystroke dynamics through the natural spontaneous use of smartphones, this study aimed to identify distinct patterns that differentiate HCs and patients with MCI.

The findings of this study revealed that keystroke dynamics, particularly HT and FT, were effective in distinguishing between HCs and patients with MCI. Notably, the sensitivity and specificity of keystroke dynamics surpassed those of the conventional screening tool. Furthermore, FT showed the highest Youden Index for all outcome measures, and its Youden Index did not increase when combined with the MoCA-K, suggesting that FT can effectively distinguish MCI on its own; this finding highlights its potential as an alternative screening method for MCI.

MCI-induced dysfunction includes motor as well as cognitive dysfunction; this finding is supported by several studies reporting that the degree of motor decline might help in dissociating MCI from healthy aging [27]. Specifically, dominant hand bradykinesia could be an efficient marker to distinguish MCI [14]. In addition, a previous study indicated that Parkinson disease signs characterized by slow movement are correlated with patients with MCI. The findings of this study, showing longer HT and FT in patients with MCI, are consistent with these studies, emphasizing the link between motor function decline and MCI [28]. Notably, motor impairment in patients with MCI considerably affects fine motor skills more than gross motor skills, which leads to a decline in dexterity skills in the upper limbs, including typing [29,30].

Declines in motor function could be attributed to working memory deficits [31,32]. Working memory is required to maintain motor chunk length in high-level motor sequence performance. However, patients with MCI, even patients with amnestic MCI, are typically characterized by deficits in working memory due to reduced brain volume in the prefrontal cortex [33]. This aligns with this study’s findings, showing lower performance in the CBT among patients with MCI compared to HCs. Interestingly, no statistically considerable difference in hand dexterity between both groups was confirmed by the grooved pegboard test in this study, suggesting that longer HT and FT in individuals with MCI are more likely to be attributed to a deficit in working memory rather than a purely motor-related issue. Therefore, smartphone typing in individuals with MCI could be negatively affected by working memory deficits.

Prior studies reported the potential of keystroke dynamics derived from touchscreen typing to capture MCI, which supports the findings of this study [13-15]. However, these studies did not compare keystroke dynamics to traditional screening tools, which limits the practical clinical applicability of keystroke dynamics [13-16]. In contrast, this study demonstrated that both HT and FT showed a greater discriminant power for MCI than the MoCA-K, suggesting that keystroke dynamics could be more beneficial to discriminate MCI from normal aging than conventional screening tools. The main factor underlying this comparative finding is based on the metrics between the keystroke dynamics and the MoCA-K. Delayed HT and FT could represent a working memory deficit, one of the hallmarks of MCI [33]. On the other hand, the MoCA-K includes orientation and language domains, which are not obviously impaired in MCI as it represents global cognitive function [5,6]. This difference in the metric could explain the superiority of the keystroke dynamics for MCI screening.

Notably, FT showed the highest Youden Index for all outcome measures, and its Youden Index did not increase when combined with the MoCA-K. This result suggests that FT is more useful in distinguishing MCI than HT. Furthermore, delayed FT can be used alone as a good marker of MCI, which indicates that a keystroke dynamics monitor could be a cost-effective way of screening MCI, considering that it does not require the additional implementation of neuropsychological tests to increase screening accuracy. Similarly, in a previous study, an ensemble model combining keystroke dynamics and linguistic features did not show superiority in discriminating MCI over keystroke dynamics alone [14], supporting the findings of this study. On the other hand, recent studies have expanded keystroke dynamics analysis by incorporating time feature analysis with graphical transformations analyzed using convolutional neural networks. Additionally, techniques such as natural language processing (eg, noun-to-verbs ratio) have been applied [15]. There have also been efforts to combine keystroke dynamics with brain imaging data analyzed through machine learning to enhance discriminant accuracy [34]. This reflects a growing interest in multimodal biomarkers of naturalistic behavioral data and machine learning methods, integrating various data sources to improve the robustness of biometric systems [35].

One of the main strengths of this study is the naturalistic collection of data in an unobtrusive way, reflecting the natural status of the participants while showing a significant correlation with the clinical screening tool. Digital health is an emerging field that enhances disease detection and management through the objective capturing of behavioral characteristics in natural situations [17]. From this digital health perspective, keystroke dynamics derived from the natural interaction with smartphones could be key factors in digital health for individuals with MCI. In addition, keystroke dynamics could be noninvasively and longitudinally captured during routine typing in real daily life without any additional effort, ensuring long-term adherence. Long adherence plays a key role in a continuous process of monitoring, diagnosis, and treatment, which is of high importance toward personalized management [14]. Considering that individualized tests that ignore the longitudinal aspect of data collection could miss people with cognitive impairment in real-world settings [36], digital biomarkers from daily smartphone use could ensure that MCI is not missed, allowing for early intervention, compared to conventional screening tools that are periodically implemented in clinics.

Despite the promising findings and strengths of this study, there are some limitations to be considered. First, these findings could not be generalized as MCI was limited to its amnestic type. Nevertheless, considering that individuals with amnestic MCI show less heterogeneity in cognitive function than those with nonamnestic MCI and multidomain MCI [5], the findings of this study have clinical implications. Second, although working memory decline in participants with MCI was investigated via the CBT, the prefrontal cortex, which underlies working memory, was not objectively observed. Thus, it is difficult to affirm whether changes in keystroke dynamics in people with MCI are due to working memory deficits caused by neurodegeneration in the prefrontal cortex. A neuroimaging study will provide more objective evidence of changes in the keystroke dynamics of patients with MCI, allowing for subgroup analysis based on neurodegeneration severity. Third, although this study confirmed through subjective reports that participants did not allow others to use their smartphones, this was not objectively verified. Finally, this study demanded technology familiarization with a smartphone to some extent.

In the future, therefore, to develop an ecologically valid MCI screening variable for a larger population, it is necessary to investigate activities that are easier to use and more widely available. Additionally, it is necessary to determine if different MCI subtypes exhibit distinct features of keystroke dynamics and incorporate these findings into brain imaging data to establish multimodal biomarkers.

Keystroke dynamics reflecting both motor and cognitive deficits were identified as a more clinically useful digital biomarker for MCI, compared to a neuropsychological screening tool. The findings of this study present a new perspective on detecting MCI via keystroke dynamics from smartphone use. These promising results suggest that measures from smartphone typing could serve as an ecologically valid digital biomarker in a nonclinical setting, acting as a surrogate for laboratory-based neuropsychological screening tools.