Andrea Bandini

Communication is a human right; despite this, children with cerebral palsy and dysarthria experience constraints due to their physical disability, affecting their opportunity to reach their full potential and ability to fully participate in play, school, and socializing. For many, their challenge to produce clear speech that can be understood is interpreted by others as an intellectual impairment. Children with moderate-severe speech impairments such as dysarthria often rely on technology solutions, which can translate text or symbols to speech manually or using a single binary switch, scanning hierarchical switch‑access menus. This is time-consuming, arduous, and 15 to 20 times slower than normal speech. Technology innovations, developed via high-quality datasets of audio-video samples of dysarthric speech, hold the key to build personalized speech-recognition algorithms that allow us to bridge this gap. My Voice Library is a repository of high-quality audio and visual datasets of children living in Australia with cerebral palsy who have dysarthria. The children are presented with gamified modules based on words and sentences from the Frenchay Dysarthria Assessment 2 to allow speech and language therapists, engineers. and programmers to develop personalized speech-recognition approaches to enable personalized real-time communication. Our ethics and consenting process enables participants (8-18 y) to choose if they want their information to be used for personalized research in addition to general research. The study has been approved by the University of Sydney's Human Research Ethics Committee (2023/263). Caregivers of the participants provide written informed consent. The results will be disseminated through contributions to international conferences and scientific journals, and they will also be included in students' theses. The National Health and Medical Research Council funded the study in April 2021. Data collection commenced in March 2024 and will continue until 99 participants have been recruited, after which the database will remain active. As of April 2026, thirteen of 99 participants have been recruited. Preliminary data analysis commenced in 2026. The results from the preliminary analysis are expected to be published in 2027, with full results to be published once 99 participants have been successfully recruited. My Voice Library embeds high-quality voice and video data assessments through gamified modules based on the Frenchay Dysarthria Assessment 2 and was designed in partnership with an active patient and a public involvement advisory group. My Voice Library and its associated information governance will facilitate the research use of real-world data where ethical approval and consent are given. My Voice Library enables the collection of voice data in controlled yet naturalistic environments for children, enhancing the clinical relevance and quality of the dataset. The modular and gamified design promotes sustained participant engagement, reduces fatigue, and increases the quantity and diversity of data collected.

Neurological disorders, such as Parkinson's disease and stroke, often lead to neurogenic dysphagia, a swallowing disorder that compromises nutrition and increases the risk of malnutrition, aspiration of food, and even death. Although instrumental assessments remain the gold standard for evaluating swallowing function, their invasiveness, costs, and reliance on specialized clinical facilities limit their routine use. In this scoping review, we provide an overview of emerging noninvasive wearable devices for assessing and monitoring dysphagia in neurological disorders. We highlight the most widely used sensing technologies, such as accelerometers, microphones, and surface electromyography, alone and in multimodal configurations, discuss their respective strengths and limitations, and outline the main challenges and opportunities for advancing this field. Particular attention is given to the need for standardized protocols, robust clinical validation, and the integration of artificial intelligence to enable scalable and precise assessment. We conclude by discussing how these technologies can be applied in clinical practice to support earlier diagnosis, continuous monitoring, and improved management of swallowing function in individuals with neurological disorders.

Electroencephalography (EEG) has shown promise in assessing and monitoring functional recovery in stroke survivors, but its utility in predicting upper limb motor recovery in a data-driven framework remains underexplored. This study presents a novel EEG-based machine-learning model, StrokeRecovNet, developed to predict motor recovery outcomes based on the upper extremity subscale of the Fugl-Meyer Assessment (FMAUE). StrokeRecovNet is a feed-forward neural network optimized for regression tasks, leveraging 221 candidate EEG biomarkers, spanning spectral and functional connectivity domains, along with baseline clinical information. These inputs are used to predict follow-up FMAUE scores in stroke survivors who underwent standard rehabilitative protocols. We validated our pipeline on two independent datasets of patients in the acute and subacute post-stroke phases. StrokeRecovNet consistently outperformed the proportional recovery rule (PRR), a standard benchmark based on initial impairment, in predicting FMAUE scores in the subacute stage (median absolute error, MAE: StrokeRecovNet = 5.85, PRR = 19.00). Incorporating support data from the subacute dataset led to improved predictive performance in the acute sample (MAE: StrokeRecovNet = 5.87, PRR = 8.80), whereas the model trained solely on the acute data did not (MAE: 13.74). Key features contributing to the model's success included brain symmetry indices and functional connectivity measures, evolving across recovery stages. These findings demonstrate the potential of EEG-based biomarkers to predict individual recovery trajectories. This work introduces a novel, data-driven approach to forecasting upper limb recovery using EEG and suggests that EEG data from the subacute stage, which is more readily available in clinical settings, can enhance early predictions, paving the way for personalized post-stroke rehabilitation strategies.

Background: Early and accurate identification of cardiac amyloidosis improves patient outcomes, yet relevant evidence is frequently hidden in free-text records. This study assesses whether structured variable extraction or direct free-text analysis more reliably identifies patients with cardiac amyloidosis, with the goal of informing clinical decision support strategies. Methods: We extracted 21 clinical variables from 432 Italian patient records using supervised and prompt-based methods with both proprietary and locally-deployable computational models. Classification performance was evaluated by comparing extracted data with gold-standard manual annotations. Two feature sets were tested: general clinical variables and amyloidosis-specific risk factors. Additionally, we evaluated direct zero-shot prediction on unstructured clinical notes. Results: For entity extraction, GPT-4.1-mini achieved F1=0.96, comparable to supervised SpaCy (F1=0.95) and GPT-4o (F1=0.94). Local open-source models like Qwen2.5 reached F1=0.94. For cardiac amyloidosis classification, machine learning models using full extracted features matched gold-standard results (SauerkrautLM-Gemma: F1=0.80 vs. gold: F1=0.82). General clinical features alone yielded lower performance (F1=0.68), highlighting that amyloidosis-specific risk factors in unstructured text provide discriminative diagnostic value. Zero-shot direct predictions outperformed supervised feature-based approaches (MedGemma: F1=0.92). Conclusions: Automated extraction and zero-shot prediction effectively structure Italian EHRs and identify amyloidosis patients without manual annotation. Domain-specific risk factors in free-text notes provide substantial predictive value. Italian hospitals can potentially deploy locally-deployable models to screen cardiac amyloidosis without manual annotation or proprietary APIs, enabling privacy-preserving clinical decision support in real-world settings.

Objective: Named Entity Recognition (NER) and Biomedical Entity Linking (BEL) are essential for transforming unstructured Electronic Health Records (EHRs) into structured information. However, tools for these tasks are limited in non-English biomedical texts such as Dutch and Italian. This study investigates the use of prompt-based learning with Large Language Models (LLMs) to perform multilingual NER and BEL using minimal domainspecific data, while addressing annotation preservation during corpus translation. Methods: An English-annotated corpus from the ShARe/CLEF dataset was translated into Dutch and Italian using a strategy that embeds annotations directly into the text prior to translation and retrieves them afterwards. GPT-4o was applied in zero-shot and few-shot settings to extract biomedical entities, which were then mapped to Unified Medical Language System Concept Unique Identifiers using contextual word embeddings. Performance was evaluated with precision, recall, and F1-score, and compared with goldstandard clinician annotations. Results: The multilingual NER pipeline achieved strong performance, with an overall F1-score of 0.98 across languages. BEL experiments showed reliable entity normalization, with an overall accuracy of 0.91 and a mean reciprocal rank of 0.95. The combined performance of the NER and BEL achieved 0.90 supporting the utility of LLMs in standardizing biomedical concepts across languages. Conclusion: Prompt-based LLMs can effectively perform NER and BEL in languages with less annotated resources, even with limited annotated training data. The proposed annotation-preserving translation method, combined with generative and discriminative LLM capabilities, provides a scalable approach to multilingual clinical information extraction. These findings highlight the potential for broader adoption of LLM-based natural language processing systems to support multilingual healthcare data harmonization. Graphical Highlights: This study shows the feasibility of using prompt-based learning with large language models (LLMs) to perform multilingual named entity recognition (NER) and biomedical entity linking (BEL) in Dutch and Italian, two languages with less annotated resources. An annotation-preserving translation strategy was proposed to adapt the ShARe/CLEF eHealth corpus, enabling consistent evaluation across English, Dutch, and Italian without loss of gold-standard annotations. The multilingual NER pipeline achieved strong overall performance (F1-score: 0.89), while BEL experiments showed reliable entity normalization (F1-score: 0.64, MRR: 0.68) to standardized clinical concepts. The approach highlights the potential of generative and discriminative LLM capabilities for scalable multilingual clinical information extraction, supporting broader European initiatives for cross-lingual health data harmonization.

Purpose: Natural Language Processing (NLP) has the potential to extract structured clinical knowledge from unstructured Electronic Health Records (EHRs). However, the limited availability of annotated datasets for algorithm training restricts its application in clinical practice. This study investigates the use of transformer-based NLP models to structure Italian EHRs in cardiac settings, addressing this gap. Methods: We implemented and evaluated three named entity recognition algorithms: SpaCy, Flair, and Multiconer. The experiments utilized three datasets comprising 2235 anamneses from patients at the Fondazione Toscana Gabriele Monasterio, Italy. Results: The SpaCy model achieved the highest performance with an F1-score of 97% in identifying clinical features on explicitly mentioned entities (Presence/Absence classification). However, features are not always mentioned, as clinicians selectively document only clinically relevant information in real-world practice. External validation shows model generalizability: EVD-100 dataset (considering 12 features, 97.13% F1) and STEMI dataset (considering 3 shared features, 88.29% F1). These structured variables were subsequently used to train machine learning algorithms (Logistic Regression, XGBoost, CatBoost) for classifying amyloidosis in heart failure patients. The classifiers trained on SpaCy-structured data attained an average F1-score of 66.70%, closely matching the 66.99% F1-score from classifiers using clinician-annotated data. Conclusion: This study shows the feasibility of using NLP for structuring Italian EHRs in realistic clinical settings, highlighting its potential to enhance computer-assisted detection despite selective documentation patterns. The comparable performance across annotation methods suggests NLP’s capability to bridge the gap in dataset annotation, paving the way for its integration into clinical practice.

Stroke remains a leading cause of long-term disability, underscoring the urgent need for effective predictors of motor recovery. Understanding the electrophysiological changes underlying spontaneous recovery could offer critical insight into recovery mechanisms and aid in predicting individual rehabilitation trajectories. In this study, we investigated the predictive power of local field potentials recorded 2 days post-stroke to forecast 1 month motor recovery in a mouse model of ischemic stroke. By employing a comprehensive machine learning approach, we identified key electrophysiological features that significantly enhanced prediction accuracy. Through nested leave-one-animal-out cross-validation, we achieved high prediction accuracy, correctly identifying motor recovery status in 15 out of 16 mice. Our findings also revealed that pre-stroke brain activity did not contribute to prediction accuracy, suggesting that post-stroke dynamics are the primary determinants of recovery. Notably, we found that features from the contralesional hemisphere were particularly influential in predicting recovery outcomes, underscoring the critical role of the non-lesioned hemisphere in motor recovery. Our data-driven methodology underscores the importance of balancing feature selection to optimize predictive performance, particularly in the context of spontaneous recovery, where insight into natural recovery processes can guide the development of targeted rehabilitation strategies. Ultimately, our findings advocate for a deeper understanding of post-stroke brain dynamics to improve clinical outcomes for stroke patients.