Artificial Intelligence in prehospital stroke detection: Automation in motion

16530
The NeuroView founders: Andrew Southerland, Omar Uribe, and Mark McDonald are pictured from left to right

Acknowledging the current limitations that hinder optimal stroke screening in the prehospital setting, Mark McDonald (University of Virginia, Charlottesville, USA), one of the co-founders of NeuroView Diagnostics—a medical technology start-up company—outlines the efforts that are underway to overcome such shortcomings. With an emphasis on artificial intelligence (AI), he details the development of new technologies, and what the future holds for them.

Stroke is the second leading cause of death and third most common cause of disability worldwide with approximately 15 million people suffering a stroke every year.1,2 Reduction in time to treatment by as little as fifteen minutes significantly decreases death and disability in stroke patients.3 Unfortunately, identifying stroke in the field is challenging. Without the use of a diagnostic aid or specialised protocol, emergency medical service (EMS) providers fail to detect stroke in as many as 40% of patients.4

Although exam based screening tools have been shown to improve stroke detection and remain the current standard of care,5 EMS providers fail to recognise roughly 15% of strokes.6,7 This figure equates to over 100,000 missed strokes in the USA if all incident strokes were screened by EMS providers.8 Many of these patients are inappropriately triaged, have a delay in treatment, or may miss the treatment window altogether. This is particularly problematic for patients with large vessel occlusion, who require urgent transport to highly specialised stroke centres for endovascular therapy. The major limitation with exam-based screening tools is that they depend on accurate interpretation of the neurological exam, which can be challenging for non-neurologists.

Efforts are currently underway to develop and implement new technologies to overcome the limitations in traditional prehospital stroke screening. One such alternative approach is to not rely on the neurological examination at all. New show promise in their ability to directly detect stroke with large vessel occlusion by assessing features of brain tissue composition or blood flow.9,10 However, it is unclear how well these devices work for detecting small lacunar or brainstem strokes. Another alternative to traditional stroke screening is the use of telemedicine. Instead of the EMS provider making the final decision about the presence of stroke, they can engage a tele-neurologist to evaluate the patient. One limitation to this approach is that it still requires the EMS provider to make a decision about the likelihood of stroke, in order to know when to engage the tele-neurologist. NeuroView seeks to improve the recognition of stroke in the prehospital setting without imaging the brain or engaging a tele-neurologist. Using artificial intelligence, we aim to automate the detection of stroke deficits in order to make better predictions about stroke in the field.

Facial weakness is a core component of the most commonly used prehospital screening instruments.11 One study demonstrated that EMS providers failed to identify weakness in 15% of stroke patients and interpreted facial weakness as being present when it was absent in 33% of patients.12 Researchers recently showed that a 3D depth camera was able to identify abnormal orofacial movements and detect stroke with 87% accuracy.13 In light of this, our team set out to automate the detection of pathological facial weakness using standard video—video that does not require the use of specialised equipment. Using computer vision and machine learning, two types of AI, our team developed an algorithm that can recognise facial weakness in standard video with 89% accuracy.14 However, our algorithm’s accuracy did not significantly differ from the average accuracy of EMS providers in our study.14 We are continually working to improve our algorithm’s performance by adding more training data, refining feature extraction, and exploring the use of different machine learning classifiers.

Automating the identification of facial weakness is an important first step in the detection of stroke in the field. In order to further refine this algorithm and develop other deficit detecting algorithms, we are currently building a video library of healthy controls and stroke patients with deficits, including facial weakness, abnormal eyes movements such as gaze deviation and nystagmus, limb weakness, and limb incoordination or ataxia. We then plan to integrate these individual deficit detecting algorithms into a stroke prediction model that we will train using video data of patients being evaluated for acute stroke. Some of these patients will have stroke; others will not. Our goal is to create AI-based software that can analyse video of a patient recorded on a mobile device and screen for stroke. Better recognition of stroke in the prehospital setting will lead to earlier and more frequent treatment with acute stroke therapy, reducing disability for thousands of stroke patients in the USA and abroad.

Beyond stroke, our mission at NeuroView is to empower non-neurologists to deliver better care when neurological expertise is unavailable. Demand for neurology in the USA exceeds supply by 11%, with a projected shortfall of 19% by 2025.15 This disparity is more pronounced worldwide with 71 of the 84 countries surveyed reporting either no neurologist or less than one neurologist per million people.16 We are starting with the misdiagnosis of stroke in the prehospital setting, but the potential impact of AI based assessment tools for neurological disease is much greater.

References

  1. Feigin, V. L., et al. Global burden of stroke. Circulation research, 120(3) (2007), 439-448.
  2. “Stroke statistics”, The internet stroke center, http://www.strokecenter.org/patients/about-stroke/stroke-statistics/
  3. Saver, Jeffrey L., et al. “Time to treatment with intravenous tissue plasminogen activator and outcome from acute ischemic stroke.” Jama 309.23 (2013): 2480-2488
  4. Smith, W. S., Isaacs, M., & Corry, M. D. (1998). Accuracy of paramedic identification of stroke and transient ischemic attack in the field. Prehospital Emergency Care, 2(3), 170-175.
  5. Adams, H. P., et al. . Guidelines for the early management of adults with ischemic stroke. Circulation, 115(20) (2007), e478-e534.
  6. Purrucker, J. C., Hametner, C., Engelbrecht, A., Bruckner, T., Popp, E., & Poli, S. (2015). Comparison of stroke recognition and stroke severity scores for stroke detection in a single cohort. J Neurol Neurosurg Psychiatry, 86(9), 1021-1028.
  7. Wojner-Alexandrov, A. W., Alexandrov, A. V., Rodriguez, D., Persse, D., & Grotta, J. C. (2005). Houston paramedic and emergency stroke treatment and outcomes study (HoPSTO). Stroke, 36(7), 1512-1518.
  8. Stroke Facts. Center for Disease Control https://www.cdc.gov/stroke/facts.htm Accessed Jul 2017
  9. Kellner, C. P., Sauvageau, E., Snyder, K. V., Fargen, K. M., Arthur, A. S., Turner, R. D., & Alexandrov, A. V. (2018). The VITAL study and overall pooled analysis with the VIPS non-invasive stroke detection device. Journal of neurointerventional surgery, 10(11), 1079-1084.
  10. 43 European Stroke Conference. 26th Conference, Berlin, Germany, 24-26 May 2017: Abstract e-Book. Cerebrovasc Dis 2017;43.
  11. Brandler, E. S., Sharma, M., Sinert, R. H., & Levine, S. R. (2014). Prehospital stroke scales in urban environments: a systematic review. Neurology, 82(24), 2241-2249.
  12. Nor, A. M., McAllister, C., Louw, S. J., Dyker, A. G., Davis, M., Jenkinson, D., & Ford, G. A. (2004). Agreement between ambulance paramedic-and physician-recorded neurological signs
  13. Bandini, A., Green, J., Richburg, B., & Yunusova, Y. (2018). Automatic Detection of Orofacial Impairment in Stroke. Proc. Interspeech 2018, 1711-1715.
  14. McDonald, M., Uribe, O., Zhuang, Y. et al (2019). Comparison of human and machine learning based facial weakness detection. Poster presented at the International Stroke Conference. Honolulu, HI.
  15. Dall, T. M., Storm, M. V., Chakrabarti, R., Drogan, O., Keran, C. M., Donofrio, P. D., … & Vidic, T. R. (2013). Supply and demand analysis of the current and future US neurology workforce. Neurology, 81(5), 470-478.
  16. Bergen, D. C., & World Federation of Neurology Task Force on Neurological Services. (2002). Training and distribution of neurologists worldwide. Journal of the neurological sciences, 198(1), 3-7.

 


LEAVE A REPLY

Please enter your comment!
Please enter your name here