Stroke is a leading cause of long-term disability worldwide. With the advancements in sensor technologies and data availability, artificial intelligence (AI) holds the promise of improving the amount, quality and efficiency of care and enhancing the precision of stroke rehabilitation. We aimed to identify and characterize the existing research on AI applications in stroke recovery and rehabilitation of adults, including categories of application and progression of technologies over time. Data were collected from peer-reviewed articles across various electronic databases up to January 2024. Insights were extracted using AI-enhanced multi-method, data-driven techniques, including clustering of themes and topics. This scoping review summarizes outcomes from 704 studies. Four common themes (impairment, assisted intervention, prediction and imaging, and neuroscience) were identified, in which time-linked patterns emerged. The impairment theme revealed a focus on motor function, gait and mobility, while the assisted intervention theme included applications of robotic and brain–computer interface (BCI) techniques. AI applications progressed over time, starting from conceptualization and then expanding to a broader range of techniques in supervised learning, artificial neural networks (ANN), natural language processing (NLP) and more. Applications focused on upper limb rehabilitation were reviewed in more detail, with machine learning (ML), deep learning techniques and sensors such as inertial measurement units (IMU) used for upper limb and functional movement analysis. AI applications have potential to facilitate tailored therapeutic delivery, thereby contributing to the optimization of rehabilitation outcomes and promoting sustained recovery from rehabilitation to real-world settings.

Altered somatosensory function is common among stroke survivors, yet is often poorly characterized. Methods of profiling somatosensation that illustrate the variability in impairment within and across different modalities remain limited. We aimed to characterize post-stroke somatosensation profiles (“fingerprints”) of the upper limb using an unsupervised machine learning cluster analysis to capture hidden relationships between measures of touch, proprioception, and haptic object recognition. Raw data were pooled from six studies where multiple quantitative measures of upper limb somatosensation were collected from stroke survivors (n = 207) using the Tactile Discrimination Test (TDT), Wrist Position Sense Test (WPST) and functional Tactile Object Recognition Test (fTORT) on the contralesional and ipsilesional upper limbs. The Growing Self Organizing Map (GSOM) unsupervised machine learning algorithm was used to generate a topology-preserving two-dimensional mapping of the pooled data and then separate it into clusters. Signature profiles of somatosensory impairment across two modalities (TDT and WPST; n = 203) and three modalities (TDT, WPST, and fTORT; n = 141) were characterized for both hands. Distinct impairment subgroups were identified. The influence of background and clinical variables was also modelled. The study provided evidence of the utility of unsupervised cluster analysis that can profile stroke survivor signatures of somatosensory impairment, which may inform improved diagnosis and characterization of impairment patterns.

Use of the upper limb to engage in everyday activities is a key indicator of functional recovery of stroke survivors. In addition to functional capacity, personal and environmental factors contribute to real-world upper limb use post-stroke. We aimed to combine data from the experience sampling method (ESM), a method used to capture real-time engagement in daily activities, with accelerometry, an objective measurement of arm use, to evaluate arm use behaviours of adult stroke survivors living in real-world environments. Thirty mild–moderately impaired stroke survivors and 30 age-standardized healthy individuals were monitored over 7 days, using accelerometers on both wrists and four ESM beeps per day to capture individual experiences in daily activities. Stroke survivors showed significantly lower use of the affected arm across all activity domains compared to the non-dominant arm of healthy participants and reported perceived lower skill and higher challenge levels. Physical context, motor capabilities and activity type were associated with affected arm use behaviour, with greater use observed during social settings and in physically demanding tasks. These findings demonstrate that combining ESM with accelerometry provides a novel, ecologically valid framework to capture and interpret the interplay between capacity, context, and behaviour in everyday life. This approach offers opportunities to design personalized, context-aware rehabilitation strategies that promote meaningful functional reintegration after stroke.

Objective: To develop and evaluate EDBERT (Emergency Department Bidirectional Encoder Representations from Transformers), a BERT-based architecture built around a multi-input fusion network that enriches free-text triage notes with the additional context of presenting complaints and patient age to predict Emergency Department (ED) disposition, supporting early decision-making and reduced ED length of stay (LOS). Methods: A retrospective cohort of 570,143 ED presentations to the Royal Melbourne Hospital, Australia, was used. At the core of EDBERT, a fusion network integrates three complementary input sources—triage notes, presenting complaints, and patient age—through learnable weights, with the two free-text inputs encoded by customised eight-layer BERT encoder stacks pre-trained on triage text with Masked Language Modelling (MLM). The resulting architecture is also lightweight, containing 66 million parameters, substantially fewer than the 110 million of BERT-BASE. Eighty percent of the dataset was used for training, and twenty percent for testing. Performance was benchmarked against BERT-Base configurations of varying depth. Results: EDBERT achieved an accuracy of 83.26%, a macro-averaged F1-score of 81.99%, and an area under the receiver operating characteristic curve of 0.91, outperforming all single-input BERT-Base variants, including a depth-matched eight-layer model (82.08%), indicating that the gain derives from the fusion of complementary inputs combined with domain-adaptive pre-training rather than from model capacity. Conclusions: A fusion-based BERT architecture that supplements the triage narrative with presenting complaint and age context predicts ED disposition more accurately than larger single-input BERT models, while requiring substantially less memory and computation, making it practical for deployment in hospitals with limited computing resources.

Background: Post-stroke somatosensory impairments, including deficits in touch and proprioception, are common. They alter body perception and can be a hidden cause of motor and functional disability. They are often underrecognized in clinical care, highlighting the need for consensus recommendations. Aim: The Somatosensory Roundtable aimed to provide the following: (A) A framework for understanding somatosensation; (B) Expert-informed consensus recommendations for somatosensory interventions that have potential for use in clinical practice with stroke survivors; and (C) Consensus recommendations for research priorities to advance clinical practice and accelerate targeted research in somatosensory recovery and rehabilitation post-stroke. Methods: An international group of clinical, research, pre-clinical, and lived-experience experts was established. A systematic search of the literature and opinions of Somatosensory Roundtable members, clinicians, and people with lived experience of stroke were used to generate lists of somatosensory interventions and research priorities for ranking in separate consensus ranking surveys. A structured consensus and voting process was adopted, with consensus ranks established through independent, individual ranking by Somatosensory Roundtable members. Results: An agreed definition and framework for understanding somatosensation was developed. Somatosensory discrimination training was ranked highest for perceived effectiveness and feasibility, with plantar tactile/proprioceptive stimulation and hands-on mechanoreceptor stimulation ranked second and third. Research priority consensus ranks were as follows: (1) effectiveness of somatosensory interventions (highest), (2) standardized measures of impairment and impact, (3) somatosensation and recovery, (4) lived experience of sensory impairment and rehabilitation, (5) individualized therapy development, (6) knowledge translation, and (7) biomarkers. Conclusion: A framework for understanding somatosensation and consensus-based recommendations for somatosensory interventions and for research priorities are provided to advance the science and practice of evidence-based rehabilitation of somatosensation after stroke.

Objective: To develop and evaluate EDBERT (Emergency Department Bidirectional Encoder Representations from Transformers), a BERT-based architecture built around a multi-input fusion network that enriches free-text triage notes with the additional context of presenting complaints and patient age to predict Emergency Department (ED) disposition, supporting early decision-making and reduced ED length of stay (LOS). Methods: A retrospective cohort of 570,143 ED presentations to the Royal Melbourne Hospital, Australia, was used. At the core of EDBERT, a fusion network integrates three complementary input sources—triage notes, presenting complaints, and patient age—through learnable weights, with the two free-text inputs encoded by customised 8-layer BERT encoder stacks pre-trained on triage text with Masked Language Modelling (MLM). The resulting architecture is also lightweight, containing 66 million parameters, substantially fewer than the 110 million of BERT-BASE. Eighty percent of the dataset was used for training and 20% for testing. Performance was benchmarked against BERT-Base configurations of varying depth. Results: EDBERT achieved an accuracy of 83.26%, a macro-averaged F1-score of 81.99%, and an area under the receiver operating characteristic curve of 0.91, outperforming all single-input BERT-Base variants, including a depth-matched 8-layer model (82.08%), indicating that the gain derives from the fusion of complementary inputs combined with domain-adaptive pre-training rather than from model capacity. Conclusion: A fusion-based BERT architecture that supplements the triage narrative with presenting complaint and age context predicts ED disposition more accurately than larger single-input BERT models, while requiring substantially less memory and computation, making it practical for deployment in hospitals with limited computing resources.