Michael Allen

Thrombolysis has been shown to reduce disability after acute ischaemic stroke in randomised trials. Our study sought to compare the benefit from thrombolysis observed in a comprehensive national stroke registry with that expected from the clinical trials. Data from a total of 168,347 ischaemic stroke patients who attended one of 118 emergency stroke hospitals in England and Wales from 2016 to 2021 were extracted from the Sentinel Stroke National Audit Programme. We constructed a machine learning model, designed to isolate the effect of thrombolysis, using explainable machine learning (XGBoost with SHapley Additive exPlanations [SHAP]). Thrombolysis was found to be associated with a statistically significant improvement in the odds of achieving a better outcome (modified Rankin scale [mRS] threshold). Regression analysis predicted a maximum 2.5-fold improvement in odds of achieving mRS 0-1, with a decline to no treatment effect at 5 h 28 min post-onset. Our results confirm a beneficial effect of thrombolysis in a large prospective national stroke registry, and align closely with meta-analyses of clinical trials of thrombolysis both in terms of magnitude of effect and decline over time. This work also demonstrates the potential to apply explainable machine learning to observational data to assist in understanding how clinical trials are implemented in real-world settings. Thrombolysis, in practice, has the same observed benefit as in the clinical trials.

Aims/hypothesis: This study aimed to develop an accessible tool, derived using machine learning, to predict hypoglycaemia risk at the start of exercise and to provide clear, rapid risk assessment to support safer participation in exercise. Methods: Data from four diverse studies were combined, encompassing 16,430 exercise sessions from 834 participants aged 12–80 years using various insulin delivery methods. The XGBoost algorithm was used to develop two models: a comprehensive model and a simplified model for predicting hypoglycaemia during exercise. Results: The comprehensive model (406 variables) achieved a mean ROC AUC of 0.89. The simplified model, using only starting glucose, exercise duration and glucose trend arrows, achieved a comparable ROC AUC of 0.87. The simplified model performed consistently across exercise types and insulin delivery methods. In collaboration with individuals with type 1 diabetes, this model was translated into GlucoseGo, a user-friendly traffic-light heatmap displaying hypoglycaemia risk based on the three variables. Conclusions/interpretation: The GlucoseGo heatmap provides a practical, accessible tool for predicting hypoglycaemia risk immediately before exercise. It may empower individuals with type 1 diabetes to exercise more safely, reduce hypoglycaemic episodes, and increase engagement in physical activity. Graphical

Background: Stroke is a leading cause of death and disability worldwide, costing the UK approximately £26 billion annually. While lifestyle modification and preventative medications reduce risk, 30% of patients with stroke have no identifiable risk factors. Hence, there is a need to better identify individuals who are at risk of stroke, and particularly those for whom the benefit of treatment outweighs the risk. AI-Based STroke Risk fActor Classification and Treatment (ABSTRACT) is a three-phase study that looks to address this issue by (1) using artificial intelligence (AI) to predict stroke risk from routine hospital data, (2) validating models on external datasets and (3) evaluating the clinical utility of AI-guided risk classification. This paper focuses on Phase I. Aims: Phase I has four objectives: Create three separate machine learning (ML) models to predict stroke risk from routine hospital data: one for CT/MRI brain data, one for ECG/echocardiography data and one for laboratory test/previous medical history data. Perform explainability analyses to identify novel stroke risk factors. Calibrate models with real-world probabilities. Combine models into a single ensemble model. Methods: This retrospective observational cohort study will include 9155 patients with stroke and 109 581 controls from southwest England (January 2003 to November 2025). Stroke cases will be identified via the Sentinel Stroke National Audit Programme. CT/MRI, ECG, echocardiography, laboratory tests, ultrasound and medical history data will be obtained from hospital and general practice records. ML techniques will then be trained on these data to predict stroke risk and identify novel stroke risk factors. Discussion: This protocol outlines ABSTRACT Phase I’s approach to creating a multimodal stroke prediction model. We detail a data handling protocol compliant with UK ethical governance, along with our strategies for data pre-processing and model training.

Background: Outcome from large vessel occlusion stroke can be significantly improved by time-critical thrombectomy but treatment is only available in regional comprehensive stroke centres (CSCs). Many patients are first admitted to a local primary stroke centre (PSC) and require transfer to a CSC, which delays treatment and decreases the chance of a good outcome. Access to thrombectomy might be improved if eligible patients could be identified in the prehospital setting and selectively redirected to a CSC. This study is evaluating a new specialist prehospital redirection pathway intended to facilitate access to thrombectomy. Methods and analysis: This study is a multicentre cluster randomised controlled trial with included health economic and process evaluations. Clusters are ambulance stations (or teams) which are work bases for ambulance practitioners. Intervention allocated ambulance practitioners use the Specialist PrE-hospital rEDirection for ischaemic stroke thrombectomY (‘SPEEDY’) pathway which comprises initiation according to specific criteria followed by contact with CSC staff who undertake a remote assessment to select patients for direct CSC admission. Control allocated ambulance practitioners continue to provide standard care which comprises admission to a local PSC and transfer to a CSC for thrombectomy if required. A co-primary outcome of thrombectomy treatment rate and time from stroke symptom onset to thrombectomy treatment will evaluate the impact of the pathway. Secondary outcomes include key aspects of emergency care including prehospital/hospital time intervals, receipt of other treatments including thrombolysis, and performance characteristics of the pathway. A broad population of all ambulance practitioner suspected and confirmed stroke patients across participating regions is being enrolled with a consent waiver. Data about SPEEDY pathway delivery are captured onto a study case record form, but all other data are obtained from routine healthcare records. Powered on a ‘primary analysis population’ (ischaemic stroke patients with pathway initiation criteria), 894 participants will detect an 8.4% difference in rate and data from 564 thrombectomy procedures will detect a 30 minute difference in time to treatment. The full study population is estimated to be approximately 80 000. Regression modelling will be used to examine primary and secondary outcomes in several analysis populations. The economic analyses will include cost-effectiveness and cost–utility analyses, and calculation of willingness to pay at a range of accepted threshold values. The process evaluation involves semi-structured interviews with professionals and patient/family members to explore views and experiences about the SPEEDY pathway. Ethics and dissemination: This study has ethical, Health Research Authority and participating NHS Trust approvals. Dissemination of study results will include presentations at national and international conferences and events, publication in peer-reviewed journals, and plain English summaries for patient/public engagement activities. Trial registration number: NCT77453332 .

Background: Artificial intelligence (AI) in healthcare often requires large, confidential clinical datasets. However, a recent UK government survey revealed that 20–40% of the public remain sceptical of its use in health research due to concerns about data security, patient–practitioner communication and commercialisation of data. A greater understanding of public attitudes is therefore needed, particularly in the context of stroke research. In this article, we describe the patient and public involvement work undertaken for the AI-Based-Stroke-Risk-fActor-Classification-and-Treatment (ABSTRACT) project, which aims to train AI models to predict future stroke risk from the electronic health records of 1 18 736 patients. Aims: We aimed to evaluate the opinions of stroke/transient ischaemic attack (TIA) patients, caregivers and members of the public on the following themes: (1) the acceptability of using AI to predict stroke from electronic health records, (2) obtaining these data using an opt-out model of consent and (3) allowing access to this dataset from members both within and outside of the routine clinical care team. Methods: A total of 83 participants were recruited via the National Health Service social media and by approaching hospital inpatients. Participants were first provided with background information on stroke, AI in medical research and ABSTRACT’s proposed data handling protocol. A mixed methods approach was then used to explore each of the above themes using online survey, semistructured focus groups and one-to-one interviews. Results: Nearly all participants felt that it was appropriate to use patient data to train AI models to predict stroke risk and that it was acceptable to obtain these data via an opt-out model of consent. Almost all participants also agreed that data could be shared within and outside of the routine clinical care team, provided it was General Data Protection Regulation compliant and used for medical research only. Conclusion: The public and those with lived stroke/TIA experience appeared to support using deidentified medical datasets for AI-driven stroke risk prediction under an opt-out consent model. However, this is provided that the research conducted is transparent, for a clear medical purpose and adheres to strict data security measures.

Aims: To evaluate the acute glucose‐lowering effect of bouts of physical activity (PA) for hyperglycaemia in individuals with type 1 diabetes, using a within‐subject matched‐pairs causal design to approximate the control condition of no activity. Methods: Data comprised 1546 PA bouts of 10–30 min from 482 participants in the T1DEXI and T1DEXIP cohorts where glucose was >10 mmol/L. Each PA bout was matched [starting glucose, glucose rate of change, insulin on board (IOB) and glucose variability (CV)] to a matched non‐PA period within the same individual using a weighted k‐nearest neighbours algorithm (SMD p < 0.001), compared with 0.3 mmol/L ( p < 0.001) during matched non‐PA periods (mean difference: −1.9 mmol/L ( p < 0.0001)). No significant differences by age, activity type or intensity were observed. The strongest predictors of PA‐induced glucose change were (in order) glucose rate of change, starting glucose, CV, duration and IOB. A heatmap using starting glucose and glucose rate of change was developed to guide real‐time decision‐making. PA‐induced hypoglycaemia risk was very low (<2%). Conclusion/Interpretation: Using PA to lower high glucose levels is an effective and safe strategy, and when guided by CGM, it can become a personalised tool for type 1 diabetes education.

Background: Intravenous thrombolysis (IVT) and mechanical thrombectomy (MT) are well-established emergency reperfusion treatments for stroke caused by clots. Both reduce disability but effectiveness is highly time-dependent, declining in the first few hours after stroke onset. Mobile stroke units (MSUs) have been proposed as a way of improving outcomes after stroke. MSUs enable on-scene brain imaging and delivery of IVT, and can allow for better choice of destination hospital. The primary objective of the study was to model the likely effect of MSUs on clinical outcomes (the ability to live independently, modified Rankin Scale 0–2) across all of England assuming no resource restrictions during deployment. Methods: We used modelling of times to treatment and outcomes. Modelling was performed for Lower Super Output Areas (LSOAs) in England. Admission numbers were based on Hospital Episode Statistics and travel times estimated from data from Open Street Map. Outcomes were predicted based on times to IVT and MT; we report outcomes as utility or the proportion of patients able to live independently at 3-6 months after stroke. We assumed MSUs and stroke units all had the same propensity to use IVT. Results: For every 100 patients suitable for IVT or MT, there will likely be 1–3 more people who can live independently following MSU care. The benefit comes from both earlier IVT and the direct transfer of patients likely to benefit from MT to their closest MT-centre by avoiding inter-hospital transfers that would be used in usual care. If, as is likely, about 1 in 5 stroke patients are suitable candidates for IVT or MT, an MSU would need to attend approximately 250 stroke patients for every one extra independent-living outcome. If about half of the patients to whom an MSU is dispatched are actual strokes (the others being stroke mimics), an MSU would need to attend approximately 500 patients for every one extra independent-living outcome. Some areas, furthest from where MSUs are based, will receive no benefit from MSU care, whereas other areas may have up to 4 additional independent-living outcomes for every 100 patients suitable for IVT or MT. Quick MSU dispatch and fast on-scene treatment are crucial to achieving the benefit of MSUs, otherwise use of MSUs may have no overall benefit, or worse outcomes, than usual care. The above benefits do not include any other possible benefits unrelated to earlier IVT or MT. Conclusions: This study suggests that the overall benefit of MSU care if deployed across all of England is likely to be modest. Selective use of MSUs in specific areas is likely to be more effective than widespread implementation. Rapid dispatch, fast on-scene treatment of patients, and careful selection of which patients to dispatch the MSU to (by location and confidence in that person being a confirmed stroke patient), are all critical for maximising benefits from MSU care. MSUs should not be seen as an alternative to optimising day-to-day emergency stroke systems.

Background: There are a lot of advances that may affect the way treatment is delivered prehospital, including mobile stroke units and point-of-care diagnostics. These have the potential to affect populations differently and therefore affect the distribution of health outcomes. Objectives: We aimed to address the following research questions: (1) Which geographic and socioeconomic inequalities have been included when evaluating access to acute stroke treatment (including reperfusion therapies)? (2) How have the identified measures been considered/assessed/calculated? (3) We also report any methodological approaches that have been proposed that might further improve the way in which acute stroke care interventions are analysed, specified relating to inequalities. Methods: PubMed and Scopus electronic databases were searched for studies that included participants who underwent acute stroke treatment and included quantitative measures of geographic and/or socioeconomic inequalities or inequities in accessing/receiving treatment. Results: Overall, sixty-six studies were included in the review. Fifty-nine included at least one measure of geographic inequalities or inequities while thirty-six included at least one measure of socioeconomic inequalities or inequities. Twenty-eight of these studies included both a geographic and socioeconomic measure of inequalities or inequities. There were no commonalities in the methods of defining, categorising and measuring the inequalities or inequities. No study provided their definition of inequality or inequity or stated any normative judgements they had made. Conclusions: It is vital that the evaluation of programmes like acute stroke care consider impacts on inequality and inequity. Researchers and policy makers should work together to determine relevant measures of inequality/inequity and the most appropriate methods of measuring and categorising them. In addition, researchers should make it clear within their work how they are defining inequality and inequity and what (if any) normative judgements have been made.

OBJECTIVE: To evaluate the acute glucose-lowering effect of bouts of physical activity (PA) for hyperglycemia in individuals with type 1 diabetes (T1D), using a within-subject matched-pairs causal design to approximate the control condition of no activity. RESEARCH DESIGN AND METHODS: Data comprised 1546 PA bouts of 10-30 min from 482 participants in the T1DEXI and T1DEXIP cohorts where glucose was >180 mg/dL (10 mmol/L). Each PA bout was matched [starting glucose, glucose rate of change, insulin on board (IOB), and glucose variability (CV) to a matched non-PA period within the same individual using a weighted k-nearest neighbors algorithm (SMD <0.01). Primary outcome: Change in glucose from PA onset to 20 minutes post-activity. Secondary outcomes: Predictors of glucose response and rate of hypoglycemia incidence. RESULTS: PA [median 23 minutes: IQR (20, 30)] led to a mean glucose change of -40 mg/dL (-2.2 mmol/L, p<0.001), compared to -5 mg/dL (0.3 mmol/L, p<0.001) during matched non-PA periods (mean difference: -35 mg/dL[1.9 mmol/L], p<0.0001). No significant differences by age, activity type, or intensity were observed. The strongest predictors of PA-induced glucose change were (in order) glucose rate of change, starting glucose, CV, duration, and IOB. A heatmap using starting glucose and glucose rate of change was developed to guide real-time decision-making. PA-induced hypoglycemia risk was very low (<2%). CONCLUSIONS: Using PA to lower high glucose levels is an effective and safe strategy, and when guided by CGM, it can become a personalized tool for T1D education. ARTICLE HIGHLIGHTS: Why did we undertake this study?: To test whether short bouts of physical activity can safely and effectively lower high glucose levels in people with type 1 diabetes. What specific question(s) did we want to answer?: Does 10–30 minutes of physical activity, started when glucose is high, reduce glucose more than doing nothing, and what factors affect this response? What did we find?: Across 1,546 activity bouts, 23 minutes of exercise lowered glucose by 40 mg/dL (2.2 mmol/L)—eight times more than rest—with <2% hypoglycemia risk. What are the implications of our findings?: Short bouts of activity quickly and safely lower high glucose, with drop size best predicted by starting level and trend—making it easy to teach and implement using a simple heatmap. Twitter Summary: Short bouts of physical activity (10–30 min) lower glucose by 40 mg/dL (2.2 mmol/L) during hyperglycemia in people with type 1 diabetes, compared to 5 mg/dL (0.3 mmol/L) with no activity. Top predictors: starting glucose, rate of change, variability, duration, insulin on board. Hypoglycemia risk was very low (<2%). CGM-guided activity is a safe, effective, and personalisable glucose-lowering tool. #T1D #CGM #Exercise Graphical