Clinical outcome modelling of rapid dynamics in acute stroke with joint-detail, continuous, remote, body motion analysis

Stroke - still the second commonest cause of death and principal cause of adult neurological disability in the Western World - is characterised by rapid changes over time and marked variability in outcomes. A patient may improve or deteriorate over minutes, and the resultant disability may range from an obvious complete paralysis to subtle, task-dependent incoordination of a single limb. Unlike many other neurological disorders, stroke can be exquisitely sensitive to prompt and intelligently tailored treatment, rewarding innovation in the delivery of care with real-world, tangible impact on patient outcomes. Optimal treatment therefore requires both detailed characterisation of the patient's clinical picture and its pattern of change over time.

Arguably the most important aspect of the patient's clinical picture - body movement - remains remarkably poorly documented: quantified only subjectively and at infrequent intervals in the patient's clinical evolution. The combination of artificial intelligence with high-performance computing now enables automatic extraction of a patient's skeletal frame resolved down to major joints, like that of a stick-man, to be delivered simply, safely, and inexpensively, without the use of cumbersome body worn markers. Central to this technology is patient privacy, with the skeletal frame extracted in real time, ensuring no video data, from which patients can be identified, to be stored or transmitted by the device.

Here we propose to use MoCat, our prototype motion categorisation system, to study the rapid dynamics of acute stroke, seamlessly embedded in the clinical stream. It can robustly determine the skeletal frame despite variations in patient size, clothing or presence of bed covering, and continuously monitor body motion at a short distance from the patient without need for extraneous wires or cables. Consequently, each bed on the hyperacute stroke unit will have its own dedicated device installed that will not interfere with the day-to-day activities on the ward. By quantifying the change in motor deficit over time we shall examine the relationship between these trajectories with clinical outcomes and develop predictive models that can support clinical management and optimise service delivery.

Past work has shown the pattern of injury to the brain from a stroke impacts on the future outcome of the patient. We will therefore create models that combine our new body motion measures, with brain scans obtained routinely as part of the hyper acute stroke pathway. In this way we not only aim to improve the accuracy of our predictions, but also examine the relationship between the pattern of stroke brain injury and motor recovery.

This project is aligned with a large-scale, Wellcome-funded, collaborative programme of translational research with the aim of creating a foundational framework for complex modelling of clinical and imaging data to predict outcomes in acute stroke.

Stroke is characterised by rapid temporal dynamics and marked functional heterogeneity, with early changes in the dynamics of upper limb recovery shown to be predictive of longer-term functional outcomes. Optimal stroke care therefore requires both detailed characterisation of the functional deficit and a high-resolution index of its rate of change. While such high-definition clinical supervision could be delivered, it would require close to a 1:1 staff/patient ratio, and could not be standardised with formal, objective models of the relation between outcomes and clinical features.

Recent advances in deep learning-assisted modelling of ordinary video data have enabled the real-world extraction of detailed body motion information, resolved down to major joints, without body-worn devices and robust to large variations in body characteristics, dress and environmental context. My clinical and engineering collaborators have developed an inexpensive miniaturised device-MoCat-that combines in analysing body motion the richness and anatomical detail of a clinician with the objectivity and automatic deployability of a machine, in real time, within real-world clinical environments. Crucially, MoCat does not store or transmit video data, preserving patient privacy, and does not require any alteration to current care pathways (e.g. addition of body markers) or major changes to a hospital's digital or estate infrastructure, rendering it readily deployable throughout the NHS and beyond.

We propose to deploy MoCat on the Hyperacute Stroke Unit at King's College Hospital. We shall quantify (1) within high-dimensional predictive models, the relationship between trajectories of motor deficits at different time scales and clinical outcomes of interest and actionable care pathway deviations; and (2) within high-dimensional lesion-deficit anatomical inferential models, the relation between trajectories of motor deficits at different temporal scales and stroke lesion patterns.

A successfully delivered project has the potential to revolutionise the management of patients with acute stroke, rendering objectively measurable the rich and rapidly-varying clinical features of this critical in-patient population, while illuminating the fundamental neuroanatomical mechanisms of rapid behavioural dynamics following focal injury to the brain. While our focus is on measures of immediate and major clinical and operational relevance in stroke care, access to an entirely new class of clinical parameter - the dynamics of continuous, joint-resolved motion - opens a myriad of possibilities across hyperacute stroke care to long-term rehabilitation. The research proposal has the potential to benefit patients, therapists, clinicians, academics, and service managers.

The system will enable continuous monitoring of the dynamics of motor dysfunction in acute stroke, facilitating better disease progression modelling. Understanding stroke dynamics will enable us accurately to estimate the course of recovery, facilitating the development of alerting systems that both detect new clinical events and identify deviations from expected patterns of clinical evolution. If adding motion information to predictive models achieves better clinical outcome predictions, we shall enhance patient stratification, enabling us to anticipate the management needs of patients, for example identifying those that could benefit from early discharge from hospital. Identifying the anatomical correlates of rapid dynamics in motor function in stroke will contribute to our understanding of the motor brain, not just in the context of stroke, but more broadly in relation to functional plasticity and neural redundancy. Clinically embedded motion capture naturally facilitates the development of interactive, game-like rehabilitation paradigms that can automatically adapt to the changing needs of the patient.

At a broader level, the research will generate significant new data which will be of interest to those involved in kinematics and modelling human motion. Our models can be used to inform future work exploring motor recovery in other disease process, such as traumatic brain injury, through a process known as transfer learning.

Collaboration with physiotherapists and occupational therapists will inform the development process, help guide the clinical relevance to short and long-term rehabilitation and facilitate the future adoption of these new measures into routine clinical practice. A presentation to the clinical therapists has been held to introduce the project and foster future engagement, with a series of future discussions planned to facilitate feedback on end-user experience.

The development of new joint motion features requires specialist knowledge regarding what aspects of motion are influential to patient performance, and an understanding of how to parametrise these complex motions numerically, constrained by the limits of the available technology. By establishing links between the clinical therapists and engineers, feature development can benefit from close discussions between the two.

We will present our work at the Clinical Research Network South London Stroke Specialty Group meeting as well as at local Neurology and Medical Grand round meetings at the Trust to raise the profile of the project amongst a wider audience interested in rehabilitation and motor recovery.