Linda Greta Dui

The assessment of handwriting is fundamental for identifying difficulties, which may have long-term negative consequences. However, standard evaluation typically focuses only on the final handwritten product. For this reason, Italian guidelines recommended supporting traditional evaluation with digital tools to also analyze the handwriting process. A sensorized ink pen used on paper was employed by over 700 students, ranging from first grade in Italian primary school to third grade in lower secondary school, to execute two tasks of the BVSCO-3, the gold standard for handwriting assessment. From sensorized ink pen data, handwriting indicators in the domains of Time, Force, Smoothness, Tilt, and Frequency were extracted. These indicators were then analyzed to examine their correlation with clinical scores, to model cross-sectional trends across grades, and to identify handwriting difficulties. The correlation analysis revealed significant relationships between the indicators and the clinical score, particularly for the Time domain. A cross-sectional statistical analysis showed that the indicators follow developmental trends compatible with handwriting learning curves reported in the literature: for many indicators, a performance plateau was reached in grade 3, from both motor and processing perspectives. Lastly, binary classification models successfully distinguished subjects with handwriting difficulties (based on BVSCO-3 results) from proficient writers. The sensorized ink pen allowed uncovering relevant characteristics of children’s handwriting process, while guaranteeing ecological data acquisition conditions. Its use could pave the way for a prompt identification of handwriting difficulties in school settings, thus facilitating an efficient referral to clinical services.

Objective: Extrauterine growth restriction (EUGR) affects 30–97% of preterm infants and is associated with poor outcomes. We used machine learning (ML) to assess how clinical and nutritional factors, particularly during the transition from parenteral to enteral nutrition, influence EUGR. Study design: This retrospective observational study included 1165 patients (46% with EUGR) born below 33 weeks’ gestation or 1500 g. We developed 10 models to predict EUGR combining two sets of features (all and nutritional features only) across five subgroups of patients (all, extremely preterm, very preterm, moderately preterm, small for gestational age). Results: Model accuracy was 0.71 (F1-score = Recall = AUROC = 0.71, Precision = 0.72) with nutritional features and 0.79 (F1-score = AUROC = 0.79, Precision = 0.80, Recall = 0.79) with all features. Lower EUGR risk was linked to female sex, higher growth velocity, and lipid intake in week one. Influential factors differed by subgroup. Conclusion: ML models accurately predicted EUGR across preterm subgroups, highlighting the role of early nutritional and clinical variables.

The screening of specific learning disabilities faces many challenges, such as: (1) the lack of an educational alliance between schools, families, and clinicians; (2) the lack of quantitative data about children's difficulties and their progression over time; or (3) the inefficiency of access to care when it is truly needed. To address these issues, this work presents ESSENCE, a platform aimed at supporting schools throughout the entire process, from identifying children with difficulties to reporting cases to child neuropsychiatrists. Following an iterative co-design process with all relevant stakeholders, several system components were defined, developed, and refined with users' feedback. The final prototype was field-tested over one year by approximately 70 children, their teachers, and some clinicians. Compliance, the System Usability Scale and custom satisfaction questionnaires were used to evaluate the system. Compliance was high, with at least five sessions conducted in 80% of the weeks. 82% of children reported good usability, and 92% were satisfied with the experience. The quantitative data collected through ESSENCE enhances the process of identifying specific learning disabilities, making it more targeted and beneficial for both children in need and the healthcare system.

Dysgraphia is a neurodevelopmental disorder specific to handwriting. Classical diagnosis is based on the evaluation of speed and quality of the final handwritten text: it is therefore delayed as it is conducted only when handwriting is mastered, in addition to being highly language-dependent and not always easily accessible. This work presents a solution able to anticipate dysgraphia screening when handwriting has not been learned yet, in order to prevent negative consequences on the individuals’ academic and daily life. To quantitatively measure handwriting-related characteristics and monitor their evolution over time, we leveraged the Play-Draw-Write iPad application to collect data produced by children from the last year of kindergarten through the second year of elementary school. We developed a meta-model based on deep learning techniques (ensemble techniques and Quasi-SVM) which receives as input raw signals collected after a processing phase based on dimensionality reduction techniques (autoencoder and Time2Vec) and mathematical tools for high-level feature extraction (Procrustes Analysis). The final dysgraphia classifier can identify “at-risk” children with 84.62% Accuracy and 100% Precision more than two years earlier than current diagnostic techniques.

Dysgraphia is a learning disability that causes handwritten production below expectations. Its diagnosis is delayed until the completion of handwriting development. To allow a preventive training program, abilities not directly related to handwriting should be evaluated, and one of them is visual perception. To investigate the role of visual perception in handwriting skills, we gamified standard clinical visual perception tests to be played while wearing an eye tracker at three difficulty levels. Then, we identified children at risk of dysgraphia through the means of a handwriting speed test. Five machine learning models were constructed to predict if the child was at risk, using the CatBoost algorithm with Nested Cross-Validation, with combinations of game performance, eye-tracking, and drawing data as predictors. A total of 53 children participated in the study. The machine learning models obtained good results, particularly with game performances as predictors (F1 score: 0.77 train, 0.71 test). SHAP explainer was used to identify the most impactful features. The game reached an excellent usability score (89.4 ± 9.6). These results are promising to suggest a new tool for dysgraphia early screening based on visual perception skills.

Handwriting learning delays should be addressed early to prevent their exacerbation and long-lasting consequences on whole children’s lives. Ideally, proper training should start even before learning how to write. This work presents a novel method to disclose potential handwriting problems, from a pre-literacy stage, based on drawings instead of words production analysis. Two hundred forty-one kindergartners drew on a tablet, and we computed features known to be distinctive of poor handwriting from symbols drawings. We verified that abnormal features patterns reflected abnormal drawings, and found correspondence in experts’ evaluation of the potential risk of developing a learning delay in the graphical sphere. A machine learning model was able to discriminate with 0.75 sensitivity and 0.76 specificity children at risk. Finally, we explained why children were considered at risk by the algorithms to inform teachers on the specific weaknesses that need training. Thanks to this system, early intervention to train specific learning delays will be finally possible.

Brain-Computer Interface (BCI) has become an established technology to interconnect a human brain and an external device. One of the most popular protocols for BCI is based on the extraction of the so-called P300 wave from electroencephalography (EEG) recordings. P300 wave is an event-related potential with a latency of 300 ms after the onset of a rare stimulus. In this paper, we used deep learning architectures, namely convolutional neural networks (CNNs), to improve P300-based BCIs. We propose a novel BCI classifier, called P3CNET, that improved P300 classification accuracy performances of the best state-of-the-art classifier. In addition, we explored pre-processing and training choices that improved the usability of BCI systems. For the pre-processing of EEG data, we explored the optimal signal interval that would improve classification accuracies. Then, we explored the minimum number of calibration sessions to balance higher accuracy and shorter calibration time. To improve the explainability of deep learning architectures, we analyzed the saliency maps of the input EEG signal leading to a correct P300 classification, and we observed that the elimination of less informative electrode channels from the data did not result in better accuracy. All the methodologies and explorations were performed and validated on two different CNN classifiers, demonstrating the generalizability of the obtained results. Finally, we showed the advantages given by transfer learning when using the proposed novel architecture on other P300 datasets. The presented architectures and practical suggestions can be used by BCI practitioners to improve its effectiveness.