The handwriting data used in this study were obtained from the DARWIN (Diagnosis Alzheimer with Handwriting) dataset, which comprises handwriting recordings from 174 participants, including individuals with Alzheimer’s disease (AD) and cognitively healthy controls, as described by Cilia et al. [15, 16]. Participants were recruited according to clinical and cognitive assessment criteria, including standard neuropsychological tests such as the Mini Mental State Examination, Frontal Assessment Battery (FAB), and Montreal Cognitive Assessment. Healthy controls were selected to match demographic and educational characteristics and were screened to exclude medications that could affect cognitive performance.

The experimental protocol includes 25 handwriting tasks performed on A4 sheets placed over a digitizing graphic tablet (Wacom Co., Ltd., Japan) which recorded pen position (x–y coordinates), pen pressure, timestamps, and in-air movements. The tasks were designed to evaluate different cognitive and motor functions and were arranged in increasing order of difficulty. The tasks include graphic drawing tasks (lines, circles, shapes), copy and reverse-copy tasks (letters, words, numbers), memory-based writing tasks, dictation tasks, functional writing tasks (e.g., postal order copying), the Clock Drawing Test, and paragraph copying. These tasks were designed to capture motor control, spatial organization, working memory, semantic processing, and motor planning abilities.

From the recorded tablet signals, stroke-level handwriting features were extracted. A stroke was defined as a continuous pen trajectory between a pen-down event and the subsequent pen-up event. From each stroke, temporal, kinematic, and pressure features were computed, including stroke duration, in-air time, velocity, acceleration, jerk, and pressure variability. These stroke-level features were subsequently aggregated and engineered to generate participant-level feature representations used for machine learning modeling.

While the original DARWIN dataset includes 25 tasks, not all tasks are equally discriminative for early Alzheimer’s detection. In this study, we focused on task-derived stroke-level features that were shown to be informative in prior work and aggregated these features from each task for subsequent analysis. The DARWIN dataset used in this study is publicly accessible via the UCI Machine Learning Repository [17].

The stroke-level data were then preprocessed, aggregated, and engineered to extract metrics such as in-air time, stroke speed, acceleration, jerk, and pressure variability. All experiments involving human subjects were conducted in accordance with the Declaration of Helsinki (2013). The original data collection received approval from the relevant institutional ethics committees, and all participants provided informed consent. Any limitations or unavailable details are explicitly acknowledged to support transparency and reproducibility. The dataset provides detailed stroke-level handwriting recordings.

For this study, sequential stroke-level features were analyzed and aggregated to capture temporal handwriting dynamics across writing sequences rather than task-specific performance alone. Stroke-level features provide higher temporal resolution and allow analysis of motor planning and execution dynamics that may not be captured by aggregated task-level handwriting measures. To ensure data quality, we conducted the following steps:

Three classifiers were evaluated:

1.

Logistic Regression (LR): Linear model for baseline comparison.

To ensure robust and unbiased evaluation, all models were assessed using five-fold stratified cross-validation. The dataset was split into five equally sized folds while preserving the proportion of AD and control participants in each fold. In each iteration, four folds were used for training and the remaining fold for testing. Performance metrics (accuracy, ROC-AUC, sensitivity, and specificity) were averaged across all folds and reported as mean values.

The overall study workflow is illustrated in Figure 1.

Model performance is summarized in Figure 2. Figure 2 compares the performance of three machine learning models LR, SVM, and RF trained on temporal and kinematic handwriting features to classify Alzheimer disease patients and healthy controls. Model evaluation was conducted using five-fold stratified cross-validation, and performance metrics (accuracy, ROC-AUC, sensitivity, and specificity) are reported as mean ± standard deviation across folds.

Across models, the Support Vector Machine achieved the highest overall performance, with the largest ROC-AUC (0.923) and highest accuracy (0.856), indicating strong discriminatory ability between Alzheimer disease patients and cognitively healthy controls. SVM also demonstrated the highest specificity (0.884), suggesting a strong ability to correctly identify healthy individuals, while maintaining competitive sensitivity (0.830).

Random Forest exhibited robust and well-balanced performance (accuracy = 0.827; ROC-AUC = 0.912; sensitivity = 0.801; specificity = 0.853) and was additionally leveraged for feature importance analysis due to its capacity to capture nonlinear relationships and feature interactions. Logistic regression showed stable but comparatively lower performance (accuracy = 0.798; ROC-AUC = 0.899; sensitivity = 0.788; specificity = 0.808), supporting its role as a transparent baseline classifier.

All performance metrics are reported as mean values across five-fold stratified cross-validation, with error bars in Figure 2 representing standard deviation, reflecting variability across folds.

Figure 3 compare baseline (4 aggregated features) and advanced (14 features including temporal, kinematic, and pressure metrics) models using five-fold stratified cross-validation. Metrics are reported as mean ± standard deviation.

As shown in Figure 3, the advanced model outperformed the baseline across all evaluation metrics:

Baseline model: accuracy = 0.799 ± 0.040, ROC-AUC = 0.887 ± 0.051

This corresponds to an approximate 4.5% absolute improvement in accuracy and a 2.7% increase in ROC-AUC, demonstrating that incorporating richer stroke-level feature representations enhances predictive performance beyond simple aggregated measures.

Analysis of feature importance using SHAP revealed that temporal dynamics of in-air movement dominated the predictive model. As shown in Figure 4, the mean and variability of in-air time, followed by pressure-related metrics, were among the most predictive features for distinguishing Alzheimer’s disease patients from healthy controls. This indicates that early motor-cognitive deficits are primarily reflected in motor planning rather than execution.

The most important features were:

1.

Mean in-air time (air_time_mean)

Visualization of stroke-level temporal patterns (Figure 5) provides direct evidence of the behavioral manifestations underlying model performance. Participants with Alzheimer’s disease exhibited systematically prolonged and more variable in-air times, along with less stable on-paper speed, across sequential strokes. These fine-grained temporal patterns are not captured by aggregated metrics, supporting the utility of temporally resolved handwriting features.

According to Figure 5, the Air time distribution (Figure 5A) demonstrates greater variability and higher median standardized values in the Alzheimer’s disease group compared to healthy, indicating consistent motor planning deficits. In the writing speed distribution (Figure 5B), Alzheimer’s disease group exhibits reduced stability and greater dispersion in on-paper speed, reflecting impaired motor execution. Pressure distribution (Figure 5C) was comparatively subtle but still contributed to group differentiation.

SHAP analysis confirmed the dominant role of in-air time dynamics in Random Forest model decision process as shown in Figure 6. Features, with higher SHAP values consistently correspond to stroke with prolonged in-air times durations and greater pressure variability, providing an interpretable link between handwriting behavior and the model’s predictions for Alzheimer’s disease.

These results provide both validation of our feature importance results and clinical interpretability for the model’s decision-making process. The primacy of in-air time dynamics provides compelling evidence that motor planning deficits precede the execution of impairments in early Alzheimer’s pathology.

To further illustrate the clinical relevance of the SHAP findings, representative feature patterns were examined. Participants in the AD group consistently exhibited elevated in-air time variability (air_time_std), increased pressure variability (pressure_mean_std), and greater stroke duration variability, indicative of irregular motor execution. In contrast, non-AD participants showed lower in-air-time dispersion, more stable pressure profiles, and consistent stroke timing. These patterns align with disruptions in motor planning, visuomotor coordination, and fine motor control reported in early Alzheimer’s disease. Mapping model predictions to observable handwriting behaviors enhances the interpretability of our framework and supports its potential utility as a non-invasive digital motor biomarker.

To place our findings in context, we compared our approach to recent studies using the DARWIN dataset and other handwriting-based AD detection methods (Table 1). While deep learning approaches such as Convolutional Neural Network, Long ShortTerm Memory networks, and hybrid transformers have achieved high accuracy and F1-scores, our study emphasizes interpretable machine learning with stroke-level temporal features. Unlike black-box deep learning models, our method provides direct insights into which handwriting dynamics contribute most to early AD detection, supporting both predictive performance and explainability.

Prolonged and variable in-air times likely reflect disruptions in motor planning mediated by prefrontal cortex and associated prefrontal-parietal networks. Preservation of on-paper speed suggests that execution processes governed by primary motor areas remain relatively intact in early AD stages. This motor-cognitive coupling deficit provides a mechanistic explanation for the observed temporal dynamics and supports handwriting as a non-invasive behavioral biomarker.

While the present study utilized the DARWIN dataset under controlled writing tasks, the results may generalize other handwriting tasks and digital devices. Future work should evaluate performance in real-world settings, including home use on tablets or smartphones.

Potential confounders such as age, education, and comorbidities may influence handwriting patterns. The DARWIN dataset primarily includes stroke-level handwriting features and class labels but does not provide detailed demographic or clinical information. As a result, these factors were not explicitly controlled in the current analysis. Future studies should incorporate such participant characteristics to ensure that observed differences in handwriting dynamics are attributable to Alzheimer’s disease rather than demographic or health-related factors.

Several limitations should be acknowledged. First, handwriting data were collected from the DARWIN dataset under standardized writing tasks, which may not fully reflect real-world writing behavior or generalize across different devices. Second, this study is cross-sectional; longitudinal follow-up is required to determine whether temporal handwriting features can reliably predict disease progression. Third, potential confounders such as age, education, and comorbidities were recorded but not explicitly controlled for. Finally, while the sample size is sufficient for exploratory analysis, generalizability across diverse populations remains to be established. Overall, the study should be considered an exploratory investigation highlighting methodological potential rather than providing definitive clinical conclusions. Overall, the study should be considered an exploratory investigation highlighting methodological potential rather than providing definitive clinical conclusions.

The integration of temporal dynamics with machine learning offers both high interpretability and practical utility. Random Forest analysis revealed key predictive features (mean and variability of in-air time and pressure), while SVM, the primary classifier, achieved high accuracy (0.856) and ROC-AUC (0.923), supporting its use for early AD detection. Future research should evaluate performance in larger, diverse cohorts, assess longitudinal prediction of disease progression, and explore integration with complementary digital biomarkers (e.g., speech, gait) to enhance early-stage detection. Evaluation across varied tasks, real-world devices, and home-based settings will be essential to ensure clinical generalizability.