Perhaps the most common reason for delays in stroke treatment can be attributed to delayed recognition of stroke symptoms by patients or families. AI solutions via wearable technologies and computer vision may help reduce this barrier by expediting EMS activation (18). Wireless sensors and wearable technologies including watches, fitness bands, and rings could be used to characterize limb motor deficits and gait patterns using accelerometers, gyroscopes, and magnetometers (18, 19). Smartphones, tablets, and computers can be turned into powerful screening tools by identifying sudden changes in motor function, speech, or language by incorporating artificial intelligence in touch screen, keyboard, and speech-to-text inputs through pressure sensors, keyboard detectors, and speech recognition software (18). This technology has the potential to alert individuals to the onset of stroke symptoms and hasten EMS activation (19, 20). For example, one recent AI algorithm reportedly reviewed 269 patients and found 97% accuracy in detection of facial asymmetry and 72% accuracy in detection of arm weakness (21, 22), and improvements in eye tracking may identify sudden changes in eye movements or unilateral gaze deviation, a dependable biomarker for LVO stroke (23, 24). In fact, observation of head and/or gaze deviation alone in telemedicine consultations has been shown to predict LVO with a sensitivity of 0.70 and specificity of 0.93, and MT necessity with a sensitivity of 0.86 and specificity of 0.90 (23). Incorporating automated detection of gaze preference, aphasia, neglect, and facial palsy into a single application may aid in early detection of acute strokes and identify patients with LVO strokes, facilitating pre-hospital triage to a comprehensive stroke center (25).

Emergency dispatcher stroke recognition is associated with higher EMS utilization of stroke scales, EMS stroke identification, and faster response times (26–28). Despite specific stroke training, the overall accuracy of emergency dispatchers leaves room for improvement (26, 29). Several groups have investigated AI as a means to improve dispatch accuracy and resource allocation. One study reviewed the potential impact of automated speech recognition software on stroke recognition in Denmark, and found that it could improve stroke detection by 16% and increase thrombolysis use by 2% (30). Another study scrutinized the application of AI algorithms in ambulance dispatch protocols for motor vehicle accidents. Through predictive analytics, this study anticipated optimal deployment times and locations for ambulances to expedite response times (31). We envision a future state where similar AI strategies can be deployed to position ambulances based on factors such as time of day, regional demographics, and other conflicting emergency calls to ensure fast response.

Generative AI and virtual reality may soon transform medical education through the development of immersive medical simulation experiences (32, 33). Stroke only accounts for approximately 2% of total EMS dispatches, and with high turnover and limited shifts, EMS personnel may lack appropriate training or experience to identify and triage stroke. While screening tools are helpful, specific simulation exercises leveraging virtual or augmented reality can provide structured educational experiences to augment training prior to real-world encounters. We envision specific software tailored to each healthcare provider role that fosters practice on simulated stroke patients in a “no-stakes” environment. These tools can provide real-time feedback to enhance retention and give EMS personnel valuable repetitions to improve performance. Similar training experiences have already been developed for surgical training and sepsis management (34, 35). Moreover, real-time AI-enhanced feedback could significantly enhance EMS provider training and competence by offering immediate insights during patient assessment, helping them to learn and adapt more effectively (30, 35, 36). This ongoing training can lead to improved performance, ensuring our EMS personnel are equipped to handle the complexities of stroke diagnosis. Nonetheless, it’s crucial to recognize that while AI can augment clinical scales and inform examinations, the inherent variability and frequency of stroke mimics leave room for error if triage decisions are solely reliant on clinical severity scores (34, 37).

Stroke neurologists were pioneering end-users of telemedicine, connecting with remote patients to provide real-time audio-video consultation in remote underserved locations for over a decade (38). The COVID-19 pandemic has accelerated the uptake of telemedicine, creating opportunities to explore AI applications to the video encounter including computer vision tools that can precisely measure human movements from videos and may allow for autonomous recognition of stroke symptoms (39). Clear use cases for image-based AI applications have already been reported in the fields such as cardiology, pathology, dermatology, and ophthalmology (40), but by applying deep-learning algorithms to live patient videos, we believe this may soon allow for the automated assessment of neurologic examinations such as the LVO stroke scales and the NIHSS. While these applications have not been tested in this clinical context, we envision the use of future computer vision applications, either alone or in conjunction with video telestroke services to more accurately recognize focal neurologic deficits and improve prehospital triage.

The ability to differentiate hemorrhagic stroke, non-LVO stroke, and LVO stroke remains a critical challenge in prehospital stroke care. AI integration with both existing and emerging technologies holds promise to develop cost-effective tools to assist in prehospital stroke diagnosis. Efforts in this arena are ongoing, including applications that harness electroencephalography (EEG) and transcranial doppler (TCD) to identify LVO. Both of these are attractive options due to their inherent portability, low-cost, and potential discriminatory ability to differentiate stroke syndromes. One study demonstrated an EEG cap during ambulance transportation can accurately differentiate LVO stroke, reporting a diagnostic accuracy of 91% and a positive likelihood ratio of 11.0 (41). Another study investigating a portable, automated TCD device and found the overall accuracy of LVO detection was 91% when combining two biomarkers. While these results demonstrated the potential value of AI to enhance existing point-of-care diagnostics, these devices have not been rigorously tested in a real-world prehospital setting yet and may not yet be ready for widespread use.

Advancements in AI have facilitated the development of novel diagnostic methods capable of swift detection and triage of stroke patients. Automated headsets have demonstrated potential in characterizing brain tissue characteristics by utilizing spectroscopy, radiofrequency, and microwave detection (42–44). AI algorithms applied in this context have been successful in differentiating patients with acute intracerebral hemorrhage and LVO stroke from healthy controls (20, 43, 44). One emerging example is a Volumetric Impedance Phase Shift Spectroscopy (VIPS) headset, a non-invasive tool that measures electrical impedance of different tissue types using electromagnetic waves. VIPS can detect lateralized changes associated with brain injury, and research suggests it can accurately identify patients with severe stroke (either LVO stroke or large ICH) with a 93% sensitivity and 87% specificity (43). Similar headsets are investigating the role of radiofrequency and low-energy microwave signals couple with machine learning algorithms to detect traumatic brain injury, intracranial hemorrhage, and LVO stroke, but results have not yet been published (44). With time, we believe these devices may provide scalable solutions to enhance existing regional stroke triage.