Exploring Age-Related Patterns in Smartphone Keystroke Dynamics Considering Temporal Variability: Cross-Sectional Study With AI-Based Analysis

Abstract

Background: Keystroke dynamics on smartphones have emerged as a promising form of passive digital biomarker. While previous studies have explored their utility in several diseases and disorders, relatively few have examined how these dynamics change systematically with chronological age in the general population.

Objective: This study aimed to investigate age-related patterns in mobile keystroke dynamics, with a particular focus on temporal variations throughout the day. By identifying behavioral signatures associated with different age groups, we further assess whether artificial intelligence–based models can accurately estimate chronological age using passively collected keystroke data.

Methods: We conducted a field study involving 177 healthy adults in the Republic of Korea, collecting free-living smartphone typing logs over multiple weeks through a custom Android keyboard app (CodeRed Corp). For each keystroke, the app recorded press and release timestamps and key type, from which 43 behavioral features were extracted across categories of speed, frequency, and temporal variability. Weekly feature vectors were constructed at 3 temporal resolutions (6-hour intervals, daily, and weekly). In total, 8 artificial intelligence models, including random forest, TabNet, transformer, and long short-term memory, were trained with participant-wise 10-fold cross-validation. A custom loss function was introduced to reduce intraparticipant prediction variability. Descriptive statistics and ablation studies were conducted to assess behavioral trends and feature contributions.

Results: The study included 177 participants (female: n=115; male: n=62) with a mean age of 28.8 (SD 11.1) years, all residing in the Republic of Korea. On average, data were collected for 25 weeks per participant, resulting in a dataset of more than 2.5 million typing sessions. Descriptive analysis revealed clear age-related differences. Younger participants typed faster and more frequently, while older participants showed slower and more variable typing. The long short-term memory model using the 6-hour interval median features achieved the best age estimation performance (mean absolute error 3.69 years, R2=0.71). When the customized loss function was applied, the model’s performance further improved to a mean absolute error of 3.60, with a reduction in intraparticipant variability in estimated ages by 7.8%. Notably, feature importance analysis suggested that the early morning (midnight to 6 AM) and late evening (6 PM to midnight) periods may carry more age-discriminative keystroke patterns.

Conclusions: Our findings demonstrated that smartphone keystroke dynamics reflect age-sensitive behavioral patterns, particularly when analyzed with fine-grained temporal resolution. While the primary goal was not age estimation per se, the ability to model these patterns highlights the potential of keystroke dynamics as a passive, unobtrusive behavioral marker for age-related functional characteristics. These insights may inform future applications in digital health, such as age-sensitive personalization or early detection of age-related decline without requiring any active user input.