We consecutively and randomly recruited 156 healthy subjects (61 males; mean age ± SD 49.6 ± 20.4 years, range 18–90 years) from the IRCCS Neuromed, Pozzilli (IS), Italy. All subjects were right-handed and native Italian speakers. We further divided participants into three age-independent subgroups to conform with previous demographic definitions of younger adults (YA) (18–35 years), middle-aged adults (MA) (36–55 years), and older adults (OA) (>56 years) (Petry, 2002). Accordingly, we examined 51 YA (23 males; 25.7 ± 3.2 years, range 18–32 years), 40 MA (17 males; 48.9 ± 5.9 years, range 37–57 years), and finally 63 OA (21 males; 71.3 ± 6.6 years, range 62–90 years). Anthropometric features including weight, height, and body mass index (BMI) were collected. The cognitive functions of all participants were assessed through the Mini-Mental State Examination (MMSE) (Folstein et al., 1975). None of the participants manifested cognitive or mood disorders, and none reported osteoarticular disorders or visual deficits significantly affecting handwriting. Also, no subjects were taking any drug affecting the central nervous system. The demographic (i.e., age and gender), anthropometric (i.e., weight, height, and BMI) and clinical (i.e., MMSE) features of the participants are reported in Table 1. All participants gave written informed consent, and the study was approved by the institutional ethics committee.

A written protocol with instructions on how to perform the handwriting task was sent to participants by the authors using the institutional email address. Also, following the enrollment, subjects received a preliminary supervised training trial to familiarize themselves with the experimental procedures. Then, participants were asked to perform the handwriting task while sitting on a chair, with their arms lying on a table, at home and in the morning. Concerning the handwriting task used in this study, participants were asked to write their own first and last names ten consecutive times on a paper sheet. We selected this specific task to exclude the influence of cultural factors and the contribution of higher cognitive abilities related to the symbolic aspects of handwriting potentially affecting the assessment of basic writing features. The first and last names were written starting from the upper and left sides of the document and then proceeding downward, in a single column, with a self-paced and comfortable speed. We preliminary provided participants with a fold of white A4 (210 mm × 297 mm) paper sheet (Fabriano, PG, Italy), and a couple of black ballpoint pen types (Bic, Clichy, France). After collecting the handwriting task three consecutive times, participants were asked to scan the signed paper sheet using the camera included in their smartphone (required resolution of at least 5 Megapixels). After scanning the handwriting samples, participants were asked to convert photos into portable document format (PDF) files using dedicated apps available for free download. Finally, participants completed the requested procedures by sending the PDF files to the authors’ institutional email server, and the files were stored anonymously on a dedicated Drive, encrypted, and password-protected (Figure 1).

All handwriting samples were preliminarily assessed using a perceptive evaluation, to exclude corrupted files from subsequent analysis. Then, based on previous data reporting changes in stroke sizes in the elderly as a prominent age-related feature (Hilton, 1977; Dixon et al., 1993; Slavin et al., 1996; Walton, 1997; Contreras-Vidal et al., 2002; Teulings et al., 2002; Rodríguez-Aranda, 2003; Rosenblum and Werner, 2006; Burger and McCluskey, 2011; Caligiuri et al., 2014), we first automatically measured the average height of strokes in the handwriting samples in each participant. As a text detection method, we used the DBNet algorithm (i.e., ResNet-50 backbone), included in the docTR package, owing to its open-source availability and state-of-the-art performance for the task (Cao et al., 2020). The DBNet, was trained according to Cao et al. (2020) and was then applied to extract the precise bounding boxes corresponding to the handwriting samples¹ (Cao et al., 2020). To limit the number of false-positive tests due to stylized strokes, we selected only bounding boxes with a height higher than 50 pixels (IC > 0.5) and relative to easily readable handwriting strokes. Finally, the DBNet algorithm calculated specific digital values as outcome measures.

The PDF files including the handwriting samples were further digitalized to fit with standard requirements for machine learning procedures, according to basic standards of artificial intelligence algorithms. A customized CNN model was built to perform classification in the age groups, following standard methodologies in the field of fine-tuning pre-trained models (Bengio et al., 2021). We classified pictures of single handwriting samples collected in the three age groups (i.e., YA, MA, and OA), according to standardized machine learning procedures, based on a fine-tuning of a pre-trained CNN with a randomly initialized fully connected layer at the end (Bengio et al., 2021). We developed a specialized multi-view CNN able to simultaneously process a set of k handwriting samples extracted from a single participant, to increase the robustness of the analysis (Figure 1). The size k of the input set was a hyper-parameter that was optimized with a dedicated grid search procedure. The architecture for further classification analysis considered the overall set as input. Accordingly, this architecture was composed of three blocks, a backbone which was used to independently process each sample in the set, a pooling block used to merge the outputs of the backbone and, finally, a classification block applied to provide a single prediction for the set (i.e., the age group of the handwriting sample). Then, after achieving a robust architecture, we organized the following steps consisting of the design of each component, the training procedure, and the hyper-parameters optimization procedures. First, the handwriting samples were resized to a single common dimension, to match the spatial dimension of the pre-trained backbone. Specifically, we resized all images to a 128 × 256-pixel proportion for all CNNs pre-trained on ImageNet (Bengio et al., 2021). To handle the handwriting samples characterized by images smaller than the optimal dimension, we applied white padding according to the respective size, matching the color of the underlying sheet of paper. We then converted the single out-of-scale image into a black and white negative image, thus obtaining large values for the corresponding output tensor. During training, we randomly selected subsets of k samples from the same PDF file as input to the network. We kept the amount of random rotation and translation to the inputs as lower as possible, to avoid image artifacts (Pang and Yang, 2016). All images were processed using a shared backbone, and we experimented with either (a) an AlexNet network trained from scratch, (b) a pre-trained ResNet-50 model, and (c) a customized Convolutional Recurrent Neural Network (CRNN), which was pre-trained on a task of optical character recognition (Wang et al., 2019). We included the choice of backbone as a second hyper-parameter to optimize the classifying procedure. We performed average pooling on the outputs of the backbone (across the spatial dimensions for the CNNs, and the temporal dimension for the CRNN), to associate each image in the input set with a fixed-dimensional embedding vector (i.e., size 1024 for the pre-trained ResNet-50 model). We applied a small fully connected layer with 100 hidden units and Gaussian Error Linear Units activation function (Hendrycks and Gimpel, 2016) on each embedding vector, and then we performed a max-pooling for all embeddings corresponding to the same input set, to obtain a fixed-dimensional vector for the entire set (Figure 1). The overall design was built to avoid the architecture being affected by the permutations of the samples in input (i.e., the output of the network does not change whether we shuffled the samples inside a set, and it could also potentially work for variable-sized sets). The final embedding vector was passed to a fully connected layer with three outputs (corresponding to the three age groups), trained with either a standard cross-entropy loss function or a weighted kappa loss function exploiting the fact that the three classes were naturally ordered (de la Torre et al., 2018; Figure 1).

To optimize the architecture, we randomly divided the dataset into three subsets with 80% of the subjects for training, 10% for validation, and 10% for test. We performed a grid search on the size of the input set, the choice of the backbone, and the selection of the loss function using the validation set, with the macro-averaged AUC score as a metric. The best performing architecture after the grid search, with a validation AUC of 81%, exploited the ResNet-50 model as the backbone, a set of size 3, and the cross-entropy loss, and we used these hyper-parameters in the following. All results shown in the experimental section were computed on the test set, by training over 50 different initializations and using the concatenation of the previously defined training and validation sets as training data. The final hyper-parameters of the model used in the experimental section were the following: the size of the crops (128; 256; RGB), number of views (Valdovinos et al., 2022), backbone model (ResNet-50—pre-trained), loss function (cross-entropy—multi-class). For the implementation, we used TensorFlow for the design of the model, with the pre-trained ResNet-50 weights taken from the official Google implementation in TensorFlow Hub, while the pre-trained CRNN weights were taken from the Keras-ocr repository.

The normal distribution of demographic and anthropometric features of subjects included in the YA, MA, and OA groups was assessed using the Kolmogorov-Smirnov test. The Chi-Square Test was used to compare the frequency of males and females in the three groups. The Mann-Whitney U-test was used to compare demographic and anthropometric parameters, as well as clinical scores (i.e., MMSE) in YA, MA, and OA. The Mann-Whitney U-test was also used to compare the stroke dimensions (i.e., heights) in participants from YA, MA, and OA groups. ROC analyses were calculated to identify the optimal cut-off values to discriminate between YA and MA, YA, and OA, and finally MA and OA, according to standardized procedures. For each ROC curve, we analyzed the Youden Index and its associated criterion (i.e., optimal threshold), as well as the Sensitivity (Se.), Specificity (Sp.), Positive Predictive Value (PPV), Negative Predictive Value (NPV), Accuracy (Acc.), Area Under the Curve (AUC), standard error (SE). Also, we compared AUCs of independent ROC curves to verify possible differences in statistical analysis, according to standardized procedures (Figure 1).

A p-value < 0.05 was considered statistically significant. Statistical analyses were conducted using STATA v17.0 (StataCorp LLC, United States).

Our first analysis demonstrated significant smaller stroke sizes in OA than in YA and MA samples, suggesting an age-related progressive decline of handwriting abilities in healthy subjects. Our findings fully agree with previous observations reporting decreased sizes along with reduced velocities and pressures of strokes during handwriting in elderly people (Hilton, 1977; Dixon et al., 1993; Slavin et al., 1996; Walton, 1997; Contreras-Vidal et al., 2002; Teulings et al., 2002; Rodríguez-Aranda, 2003; Kim et al., 2005; Rosenblum and Werner, 2006; Burger and McCluskey, 2011; Plamondon et al., 2013; Caligiuri et al., 2014; Zham et al., 2019; Kanno et al., 2020). By contrast, some reports suggested increased rather than decreased stroke size in elderly subjects (Rosenblum et al., 2013) raising the possibility of relevant heterogeneity in previous methodologies and experimental approaches. Abnormal stroke sizes observed in the elderly during handwriting would point to the effect of physiological aging on brain networks contributing to this high-level cognitive function (Purcell et al., 2011; Planton et al., 2013; Bartoň et al., 2020). Clinical, neuropsychological and neuroimaging studies have demonstrated that handwriting relies on the widespread synergic activity of several brain regions, the writing network. More in detail, the angular gyrus (Roeltgen and Heilman, 1984) and the precentral gyrus (Rapcsak et al., 1988) contribute to lexical processes of handwriting, the left perisylvian regions play a role in phonological aspects (Roeltgen and Heilman, 1984; Rapcsak et al., 1988, 2009; Alexander et al., 1992b), whereas the left superior parietal or premotor regions are responsible for handwriting execution (Auerbach and Alexander, 1981; Anderson et al., 1990; Alexander et al., 1992a). Indeed, selective brain lesions in the angular gyrus/precentral gyrus/left perisylvian regions and in the left superior parietal/premotor regions are known to lead to dysorthographias (i.e., lexical or phonological components) (Roeltgen and Heilman, 1984; Rapcsak et al., 1988, 2009; Alexander et al., 1992b) and apraxic agraphia (i.e., grapheme tracing) (Auerbach and Alexander, 1981; Anderson et al., 1990; Alexander et al., 1992a), respectively. Lastly, the writing network also reflects the activation of subcortical structures such as the basal ganglia (i.e., striatum) and the anterior cerebellum (Kim et al., 2005; Purcell et al., 2011; Planton et al., 2013, 2017; Letanneux et al., 2014; Zham et al., 2019; Bartoň et al., 2020; Kanno et al., 2020). In the elderly, the age-related progressive reduction in stroke sizes during handwriting would therefore reflect structural or functional changes in cortico-subcortical components of the writing network (Purcell et al., 2011; Planton et al., 2013; Bartoň et al., 2020). Several previous reports have demonstrated an overall reduction of default subnetwork connectivity in elderly subjects when performing experimental paradigms assessing executive functions, such as handwriting (Planton et al., 2017; Schulz et al., 2022). This would reflect several age-related biological factors including progressive white matter involvement in the frontal lobe attributable to a small vessel disease (de Leeuw et al., 2001; Kim and Melhem, 2008; Walhovd et al., 2011; Hirsiger et al., 2017; Zhuang et al., 2018). A further consideration concerns the possible link between the age-related progressive reduction in stroke sizes and the parkinsonian micrographia. Previous observations have demonstrated that micrographia in PD is related to the severity of bradykinesia (Wu et al., 2016; Wyss-Coray, 2016; Schott, 2017; Canevelli and Marsili, 2022), thus reflecting dopaminergic striatal denervation (Hilton, 1977; Dixon et al., 1993; Slavin et al., 1996; Walton, 1997; Contreras-Vidal et al., 2002; Teulings et al., 2002; Rodríguez-Aranda, 2003; Rosenblum and Werner, 2006; Burger and McCluskey, 2011; Plamondon et al., 2013; Rosenblum et al., 2013; Caligiuri et al., 2014; Wu et al., 2016; Kanno et al., 2020). Accordingly, a further hypothesis would attribute the age-related reduction in stroke sizes observed in healthy elderly to abnormal activity in cortico-subcortical components of the writing network which at least in part overlap with those responsible for micrographia in PD (Kim et al., 2005; Rosenblum et al., 2013; Wu et al., 2016; Zham et al., 2019; Kanno et al., 2020). Future studies will disclose possible similarities between age-related reduction of stroke sizes and parkinsonian “consistent” (i.e., overall small handwriting) and “progressive” micrographia (i.e., serial reduction in handwriting size) (Kim et al., 2005; Rosenblum et al., 2013; Wu et al., 2016; Zham et al., 2019; Kanno et al., 2020). Despite PD, converging clinical, neuropsychological and neuroimaging evidence point to abnormal handwriting abilities also in patients with other neurodegenerative disorders including AD (Forbes et al., 2004; Delazer et al., 2021). Patients with AD may manifest agraphia associated with a reduction in stroke sizes (Delazer et al., 2021). Overall, given that micrographia constitutes a reliable writing feature in PD (Letanneux et al., 2014) and possibly characterizes also other neurodegenerative diseases such as AD (Delazer et al., 2021), we speculate that future handwriting analysis would help to assess the phenoconversion from physiologic aging into neurodegenerative disorders (Letanneux et al., 2014).

Relevant new findings also came from our machine learning analysis which allowed us to achieve high accuracy obtained when discriminating handwriting samples collected from YA and OA. This finding, which has been achieved objectively and automatically, strongly supports the hypothesis of a relevant detrimental effect of human aging on handwriting samples. Further relevant information on age-related changes in handwriting came from the high accuracy obtained when discriminating between handwriting samples from MA and OA. Given that our cohort of MA aged from 37 to 57 years, whereas the group of OA ranged from 62 to 90 years, our results would raise the intriguing hypothesis that handwriting prominently declines at the age of about 60 years. These findings agree with a previous study based on standard handwriting analysis which has shown a consistent deterioration of handwriting skills starting from 60 years of age (Verwey et al., 2011). This would reflect a relevant cognitive workload associated with a global reduced sensory processing performance leading to a prominent deterioration of handwriting skills in the elderly of similar age ranges (Engel-Yeger et al., 2012).

A rather unexpected finding of our study consists of the high accuracy achieved by machine learning when classifying between handwriting collected in YA and MA. Given that our cohort of YA ranged from 18 to 32 years, whereas MA aged from 37 to 57 years, our results would suggest early age-related changes in handwriting skills. The differences here observed when comparing handwriting between YA and MA would reflect several possible mechanisms. A first hypothesis concerns a putative early deterioration of brain networks responsible for complex motor tasks at about 35 years of age. Such a hypothesis would receive support from the observation that since the mid-twenties, complex motor, as well as cognitive functions progressively, decline (Berghuis et al., 2019; Stojan and Voelcker-Rehage, 2021). Given that brain networks regulating handwriting partially overlap with those controlling abstract thinking or complex motor planning, we believe that handwriting deteriorates in MA as a result of the relentless phenomenon of aging (Fraser et al., 2009). An alternative hypothesis would imply non-biological mechanisms, including social issues. Our group of YA consists of participants who belong to the so-called “millennial generation” and are possibly characterized by social- and cultural-driven changes in writing habits. The “millennial generation” indeed concerns teenagers and younger adults born since 1985 and refers to people that grew up in the internet age. These youngers largely use computers and other technological devices for writing their academic essays from the early school classes and for communicating on social media. Hence, instead of traditional writing, they prefer typing on a clipboard. Accordingly, it can be argued that this recently acquired social habit would impact significantly on handwriting skills in YA. Future studies could confirm our speculation on social-related changes in handwriting patterns in the “millennial generation” (Dressler et al., 2018).

A final consideration concerns the real-life setting of our experimental design. Indeed, we here report the first handwriting analysis based on machine learning to apply homemade recordings of handwriting samples. The high significance of the results achieved in such an ecologic scenario discloses plenty of future perspectives, including social distancing and disparities in the access to care (Chirra et al., 2019). Therefore, we suggest that future telemedicine studies on handwriting analysis in physiologic and pathologic aging should be based on the examination of their impact on daily activities, to improve daily performance and quality of life (Marsili et al., 2018; Marsili and Mahajan, 2022). Lastly, our study would also be useful to develop a methodology based on the automatic machine learning analysis of handwriting according to age. For instance, an automatic technique able to recognize age-related features of signatures would be applied by authorities to discriminate against authentic wills in heritage disputes or for dating documents and paintings.

We acknowledge that our study has potential limitations. Our sample size as well as the number of handwriting samples collected in healthy subjects would be considered relatively small. However, the accuracy achieved in the classification of handwriting in the three cohorts of healthy subjects was remarkable pointing to the reliability of our machine learning analysis. Also, given that we have not recorded handwriting samples serially in each participant, our study does not allow us to reach conclusions about the intrasubject variability in handwriting skills. Furthermore, we cannot fully exclude that the decreased weight and BMI observed in YA, and the decreased height observed in OA subjects would have contributed at least in part to our findings.