AbstractBackground & MotivationStroke outcomes following revascularization therapy (intravenous thrombolysis, IVT, and/or mechanical thrombectomy, MT) depend critically on time from stroke onset to treatment. Different service configurations may prioritise time to IVT or time to MT. In order to evaluate alternative acute stroke care configurations, it is necessary to estimate clinical outcomes given differing times to IVT and MT.MethodModel using an algorithm coded in Python. This is available at https://github.com/MichaelAllen1966/stroke_outcome_algorithmResultsWe demonstrate how the code may be used to estimate clinical outcomes given varying times to IVT and MT.ConclusionPython code has been developed and shared to enable estimation of clinical outcome given times to IVT and MT. Here we share pseudocode and links to full Python code.

Objectives: to guide policy when planning reperfusion thrombolysis (IVT) and thrombectomy (MT) services for acute stroke in England, focussing on the choice between ‘mothership’ (direct conveyance to an MT centre) and ‘drip-and-ship' (secondary transfer for MT after local IVT) provision and the impact of bypassing local acute stroke centres.Methods: Computer modelling was used to estimate the likely outcomes from reperfusion therapies, along with admission numbers to units, based on expected times to IVT and MT.Results: Without pre-hospital selection for LAO, 94% of the population of England live in areas where the greatest clinical benefit accrues from direct conveyance to an IVT/MT centre. If this model was followed then net benefit from reperfusion is predicted to be increased from 31 to 34 additional disability-free outcomes / 1,000 admissions. However, this policy produces unsustainable admission numbers at these centres, and depletes all but 19 IVT-only units of all stroke admissions. Implementing a maximum permitted additional travel time to bypass an IVT-only unit, or using a pre-hospital test for LAO, both increase net benefit over the current drip-and-ship model, but produce a similar destabilising effect on acute systems of care. Use of IVT-only units manage admission numbers to IVT/MT centres.Conclusions: the mothership model reduces time to MT at the cost of increased time to IVT, but the benefit of faster MT is predicted to lead to a modest improvement in overall outcomes. Providing a sustainable national system of acute stroke care requires a hybrid of mothership and drip-and-ship provision.

Objectives: to guide policy when planning thrombolysis (IVT) and thrombectomy (MT) services for acute stroke in England, focussing on the choice between ‘mothership’ (direct conveyance to an MT centre) and ‘drip-and-ship’ (secondary transfer) provision and the impact of bypassing local acute stroke centres. Design: Outcome-based modelling study. Setting: 107 acute stroke centres in England, 24 of which provide IVT and MT (IVT/MT centres) and 83 provide only IVT (IVT-only units). Participants: 242,874 emergency admissions with acute stroke over 3 years (2015–2017). Intervention: Reperfusion delivered by drip-and-ship, mothership or ‘hybrid’ models; impact of additional travel time to directly access an IVT/MT centre by bypassing a more local IVT-only unit; effect of pre-hospital selection for large artery occlusion (LAO). Main outcome measures: Population benefit from reperfusion, time to IVT and MT, admission numbers to IVT-only units and IVT/MT centres. Results: Without pre-hospital selection for LAO, 94% of the population of England live in areas where the greatest clinical benefit, assuming unknown patient status, accrues from direct conveyance to an IVT/MT centre. However, this policy produces unsustainable admission numbers at these centres, with 78 out of 83 IVT-only units receiving fewer than 300 admissions per year (compared to 3 with drip-and-ship). Implementing a maximum permitted additional travel time to bypass an IVT-only unit, using a pre-hospital test for LAO, and selecting patients based on stroke onset time, all help to mitigate the destabilising effect but there is still some significant disruption to admission numbers, and improved selection of patients suitable for MT selectively reduces the number of patients who would receive IVT at IVT-only centres, challenging the sustainability of IVT expertise in IVT-only centres. Conclusions: Implementation of reperfusion for acute stroke based solely on achieving the maximum population benefit potentially leads to destabilisation of the emergency stroke care system. Careful planning is required to create a sustainable system, and modelling may be used to help planners maximise benefit from reperfusion while creating a sustainable emergency stroke care system.

BACKGROUND: Expert opinion is that about 20% of emergency stroke patients should receive thrombolysis. Currently, 11% to 12% of patients in England and Wales receive thrombolysis, ranging from 2% to 24% between hospitals. The aim of this study was to assess how much variation is due to differences in local patient populations, and how much is due to differences in clinical decision-making and stroke pathway performance, while estimating a realistic target thrombolysis use. METHODS: Anonymised data for 246 676 emergency stroke admissions to 132 acute hospitals in England and Wales between 2016 and 2018 was obtained from the Sentinel Stroke National Audit Programme data. We used machine learning to learn decisions on who to give thrombolysis to at each hospital. We used clinical pathway simulation to model effects of changing pathway performance. Qualitative research was used to assess clinician attitudes to these methods. Three changes were modeled: (1) arrival-to-treatment in 30 minutes, (2) proportion of patients with determined stroke onset times set to at least the national upper quartile, (3) thrombolysis decisions made based on majority vote of a benchmark set of hospitals. RESULTS: of the modeled changes, any single change was predicted to increase national thrombolysis use from 11.6% to between 12.3% to 14.5% (clinical decision-making having the most effect). Combined, these changes would be expected to increase thrombolysis to 18.3%, but there would still be significant variation between hospitals depending on local patient population. Clinicians engaged well with the modeling, but those from hospitals with lower thrombolysis use were most cautious about the methods. CONCLUSIONS: Machine learning and clinical pathway simulation may be applied at scale to national stroke audit data, allowing extended use and analysis of audit data. Stroke thrombolysis rates of at least 18% look achievable in England and Wales, but each hospital should have its own target.

BackgroundStroke is a common cause of adult disability. Expert opinion is that about 20% of patients should receive thrombolysis to break up a clot causing the stroke. Currently, 11–12% of patients in England and Wales receive this treatment, ranging between 2% and 24% between hospitals.ObjectivesWe sought to enhance the national stroke audit by providing further analysis of the key sources of inter-hospital variation to determine how a target of 20% of stroke patients receiving thrombolysis may be reached.DesignWe modelled three aspects of the thrombolysis pathway, using machine learning and clinical pathway simulation. In addition, the project had a qualitative research arm, with the objective of understanding clinicians’ attitudes to use of modelling and machine learning applied to the national stroke audit.Participants and data sourceAnonymised data were collected for 246,676 emergency stroke admissions to acute stroke teams in England and Wales between 2016 and 2018, obtained from the Sentinel Stroke National Audit Programme.ResultsUse of thrombolysis could be predicted with 85% accuracy for those patients with a chance of receiving thrombolysis (i.e. those arriving within 4 hours of stroke onset). Machine learning models allowed prediction of likely treatment choice for each patient at all hospitals. A clinical pathway simulation predicted hospital thrombolysis use with an average absolute error of 0.5 percentage points. We found that about half of the inter-hospital variation in thrombolysis use came from differences in local patient populations, and half from in-hospital processes and decision-making. Three changes were applied to all hospitals in the model: (1) arrival to treatment in 30 minutes, (2) proportion of patients with determined stroke onset times set to at least the national upper quartile and (3) thrombolysis decisions made based on majority vote of a benchmark set of 30 hospitals. Any single change alone was predicted to increase national thrombolysis use from 11.6% to between 12.3% and 14.5% (with clinical decision-making having the most effect). Combined, these changes would be expected to increase thrombolysis to 18.3% (and to double the clinical benefit of thrombolysis, as speed increases also improve clinical benefit independently of the proportion of patients receiving thrombolysis); however, there would still be significant variation between hospitals depending on local patient population. For each hospital, the effect of each change could be predicted alone or in combination. Qualitative research with 19 clinicians showed that engagement with, and trust in, the model was greatest in physicians from units with higher thrombolysis rates. Physicians also wanted to see a machine learning model predicting outcome with probability of adverse effect of thrombolysis to counter a fear that driving thrombolysis use up may cause more harm than good.LimitationsModels may be built using data available in the Sentinel Stroke National Audit Programme only. Not all factors affecting use of thrombolysis are contained in Sentinel Stroke National Audit Programme data and the model, therefore, provides information on patterns of thrombolysis use in hospitals, but is not suitable for, or intended as, a decision aid to thrombolysis.ConclusionsMachine learning and clinical pathway simulation may be applied at scale to national audit data, allowing extended use and analysis of audit data. Stroke thrombolysis rates of at least 18% look achievable in England and Wales, but each hospital should have its own target.Future workFuture studies should extend machine learning modelling to predict the patient-level outcome and probability of adverse effects of thrombolysis, and apply co-production techniques, with clinicians and other stakeholders, to communicate model outputs.FundingThis project was funded by the National Institute for Health and Care Research (NIHR) Health and Social Care Delivery Research programme and will be published in full inHealth and Social Care Delivery Research; Vol. 10, No. 31. See the NIHR Journals Library website for further project information.

Background and motivation: Deep Reinforcement Learning (Deep RL) is a rapidly
developing field. Historically most application has been made to games (such as
chess, Atari games, and go). Deep RL is now reaching the stage where it may
offer value in real world problems, including optimisation of healthcare
systems. One such problem is where to locate ambulances between calls in order
to minimise time from emergency call to ambulance on-scene. This is known as
the Ambulance Location problem.
. Aim: to develop an OpenAI Gym-compatible framework and simulation environment
for testing Deep RL agents.
. Methods: a custom ambulance dispatch simulation environment was developed
using OpenAI Gym and SimPy. Deep RL agents were built using PyTorch. The
environment is a simplification of the real world, but allows control over the
number of clusters of incident locations, number of possible dispatch
locations, number of hospitals, and creating incidents that occur at different
locations throughout each day.
. Results: a range of Deep RL agents based on Deep Q networks were tested in
this custom environment. All reduced time to respond to emergency calls
compared with random allocation to dispatch points. Bagging Noisy Duelling Deep
Q networks gave the most consistence performance. All methods had a tendency to
lose performance if trained for too long, and so agents were saved at their
optimal performance (and tested on independent simulation runs).
. Conclusions: Deep RL agents, developed using simulated environments, have the
potential to offer a novel approach to optimise the Ambulance Location problem.
Creating open simulation environments should allow more rapid progress in this
field.

OBJECTIVE: to evaluate the application of clinical pathway simulation in machine learning, using clinical audit data, in order to identify key drivers for improving use and speed of thrombolysis at individual hospitals. DESIGN: Computer simulation modelling and machine learning. SETTING: Seven acute stroke units. PARTICIPANTS: Anonymised clinical audit data for 7864 patients. RESULTS: Three factors were pivotal in governing thrombolysis use: (1) the proportion of patients with a known stroke onset time (range 44%-73%), (2) pathway speed (for patients arriving within 4 hours of onset: per-hospital median arrival-to-scan ranged from 11 to 56 min; median scan-to-thrombolysis ranged from 21 to 44 min) and (3) predisposition to use thrombolysis (thrombolysis use ranged from 31% to 52% for patients with stroke scanned with 30 min left to administer thrombolysis). A pathway simulation model could predict the potential benefit of improving individual stages of the clinical pathway speed, whereas a machine learning model could predict the benefit of 'exporting' clinical decision making from one hospital to another, while allowing for differences in patient population between hospitals. By applying pathway simulation and machine learning together, we found a realistic ceiling of 15%-25% use of thrombolysis across different hospitals and, in the seven hospitals studied, a realistic opportunity to double the number of patients with no significant disability that may be attributed to thrombolysis. CONCLUSIONS: National clinical audit may be enhanced by a combination of pathway simulation and machine learning, which best allows for an understanding of key levers for improvement in hyperacute stroke pathways, allowing for differences between local patient populations. These models, based on standard clinical audit data, may be applied at scale while providing results at individual hospital level. The models facilitate understanding of variation and levers for improvement in stroke pathways, and help set realistic targets tailored to local populations.

Purpose: Both intravenous thrombolysis (IVT) and intra-arterial endovascular thrombectomy (ET) improve the outcome of patients with acute ischaemic stroke, with endovascular thrombectomy being an option for those patients with large vessel occlusions. We sought to understand how organisation of services affects time to treatment for both intravenous thrombolysis and endovascular thrombectomy. Method: a multi-objective optimisation approach was used to explore the relationship between the number of intravenous thrombolysis and endovascular thrombectomy centres and times to treatment. The analysis is based on 238,887 emergency stroke admissions in England over 3 years (2013–2015). Results: Providing hyper-acute care only in comprehensive stroke centres (CSC, providing both intravenous thrombolysis and endovascular thrombectomy, and performing >150 endovascular thrombectomy per year, maximum 40 centres) in England would lead to 15% of patients being more than 45 min away from care, and would create centres with up to 4300 stroke admissions/year. Mixing hyper-acute stroke units (providing intravenous thrombolysis only) with comprehensive stroke centres speeds time to intravenous thrombolysis and mitigates admission numbers to comprehensive stroke centres, but at the expense of increasing time to endovascular thrombectomy. With 24 comprehensive stroke centres and all remaining current acute stroke units as hyper-acute stroke units, redirecting patients directly to attend a comprehensive stroke centre by accepting a small delay (15-min maximum) in intravenous thrombolysis reduces time to endovascular thrombectomy: 25% of all patients would be redirected from hyper-acute stroke units to a comprehensive stroke centre, with an average delay in intravenous thrombolysis of 8 min, and an average improvement in time to endovascular thrombectomy of 80 min. The balance of comprehensive stroke centre:hyper-acute stroke unit admissions would change from 24:76 to 49:51. Conclusion: Planning of hyper-acute stroke services is best achieved when considering all forms of acute care and ambulance protocol together. Times to treatment need to be considered alongside manageable and sustainable admission numbers.

Stroke care has been identified as an area where operations research has great potential. In recent years there has been a small but sustained stream of discrete-event simulation case studies in modelling hyperacute stroke systems. The nature of such case studies has led to a fragmented knowledge base and high entry cost to stroke modelling research. Two common issues have faced researchers in stroke care: understanding the logistics and clinical aspects of stroke care and moving from these findings to an appropriately detailed model. We aim to accelerate studies in this area by introducing a conceptual modelling framework that is domain specific for stroke. A domain specific framework trades-off the wide applicability of a general framework against increased efficiency and reuse to support modelling in the problem domain. This compromise is appropriate when the problem domain is complex, of high value to society, and where the saving in future modelling effort is likely to be greater than the effort to create the framework. We detail the requirements of a domain specific conceptual model and then provide domain specific knowledge to support modellers in gaining an understanding of the problem situation, translating this knowledge into selected model outputs, inputs and content in the case of hyperacute stroke. We illustrate the use of the framework with an example based at a large hospital in the United Kingdom.

OBJECTIVES: the policy of centralising hyperacute stroke units (HASUs) in England aims to provide stroke care in units that are both large enough to sustain expertise (>600 admissions/year) and dispersed enough to rapidly deliver time-critical treatments (

This paper describes a procedure to use amodel interactively to investigate future land use by studying a wide range of scenarios defining climate, technological and socio-economic changes. A full model run of several hours has been replaced by a metamodel version which takes a few seconds, and provides the user with an immediate visual output and with the ability to examine easily which factors have the greatest effect. The Regional Impact Simulator combines a model of agricultural land use choices linked with models of urban growth, flooding risk, water quality and consequences for wildlife to estimate plausible futures of agricultural land on a timescale of 20–50 years. The model examines the East Anglian and North West regions of the United Kingdom at a grid resolution of 5 × 5 km, and for each scenario estimates the most likely cropping and its profitability at each location, and classifies land use as arable, intensive or extensive grassland or abandoned. From a modelling viewpoint the metamodel approach enables iteration. It is thus possible to determine how product prices change so that production meets demand. The results of the study show that in East Anglia cropping remains quite stable over a wide range of scenarios, though grassland is eliminated in scenarios with the 2050s High climate scenario – almost certainly due to the low yield in the drier conditions. In the North West there is a very much greater range of outcomes, though all scenarios suggest a reduction in grassland with the greatest in the 2050s High climate scenario combined with the “Regional Stewardship” (environmental) socio-economic scenario. The effects of the predicted changes in land use on plant species showed suitability for species to vary greatly, particularly between the socio-economic scenarios, due to detrimental effects from increases in nitrogen fertilisation. A complete simulation with the Regional Impact Simulator takes around 15 seconds (computer-dependent), which users who responded felt was adequate or better than adequate. The main areas for future improvement, such as the speed of the system, user interaction and the accuracy and detail of the modelling, are considered.