Stroke Snapshot: Incorporating Artificial Intelligence in Stroke Management: Redefining State of the Art

Stroke is a leading cause of death and severe long-term disability worldwide.¹ The burden of stroke affects people along a continuum of care, resulting in a need for high-quality evidence and novel interventions that target prevention, prognostication, acute treatment, and recovery and rehabilitation. This is an exciting period in the stroke field, with frequent new discoveries and clinical trials on optimal management, accompanied by an impetus to make decisions rapidly and accurately based on increasingly individualized and complex information. This paradigm provides an opportunity for stroke clinicians to use artificial intelligence (AI) for complex decision-making. As new diagnostic and intervention tools become available and the relationships between individual, disease, and treatment factors are better characterized, clinicians will need to acquire and fine-tune multiple skills (eg, imaging interpretation, patient assessment, data mining) while staying up to date and integrating the most recent data into their clinical practice. The development of comprehensive predictive models based on individual pathophysiology will be critical for achieving standardized and efficient assessments to improve overall stroke management.

Defining AI for Clinical Practice

The incorporation of AI into the medical field marks the beginning of a new era that is expected to yield insights beyond the pace and extent of what could be accomplished by individuals. AI is defined broadly as the design, evaluation, and use of nonhuman intelligent systems that are capable of perceiving, synthesizing, and inferring information from their environments, ultimately rendering them capable of mimicking human cognitive tasks such as learning and problem solving.² Machine learning (ML) is a subset of AI characterized by the automated ability to learn and incorporate newly generated data or experiences without the need for additional programming, enabling AI to make systemic inferences in contrast to the reductionist approaches of traditional methods.² ML algorithms can be particularly useful for treating highly complex and heterogeneous neurovascular disorders, such as stroke. Deep learning is a subset of ML that incorporates unsupervised learning methods, such as convolutional neural networks, which allows the identification of nonlinear relationships between person-specific characteristics and stroke severity. This quality has the disadvantage of creating a “black box” between a problem and its proposed solution, whereby the user has limited ability to understand the process taken to obtain that solution. Despite this limitation, ML and deep learning methods can be used to facilitate and expedite the clinical decision-making process and assist with informed discussions with individuals and their families.^2–4 AI has substantial potential for improvements spanning the whole of stroke management: from faster diagnosis and acute treatment administration to more effective prevention and potentially more complete recovery.

Acute Stroke Management

In acute stroke care, AI is proving to be particularly useful in early stroke detection and identification of eligibility for acute reperfusion therapy (eg, intravenous thrombolytics or endovascular mechanical thrombectomy). The time-sensitive therapeutic window for acute ischemic stroke, which necessitates the rapid analysis of large amounts of data to enable complex clinical decisions to be made in a short time frame, makes accurate diagnosis with early treatment a cornerstone of effective management. Identifying key elements of a person’s medical history quickly, often through mining extensive electronic medical records, remains a rate-limiting step for determining safety and eligibility for acute stroke treatment. The potential for AI tools to sift through clinical data to provide a summary of pertinent medical history poses an exciting opportunity to reduce time to intervention while maintaining patient safety.⁵

AI tools are being developed to augment the clinical assessment of acute stroke, particularly in settings where immediate neurologic expertise is not available. This includes supporting prehospital stroke screening and triage by emergency medical services, and augmenting the neurologic examination by physicians and nurses in rural or resource-limited emergency departments. In a video-based study of people with facial weakness, an ML algorithm was able to detect unilateral facial weakness with accuracy similar to that of trained paramedics, while having better accuracy in identifying laterality (Figure 1).⁶ A deep-learning multimodal framework, DeepStroke, was able to analyze audio–video samples from real-world individuals to detect facial paralysis and global speech disorders with a higher sensitivity and accuracy than that of emergency providers.⁷

Neurovascular Imaging

Several studies have evaluated the use of AI to enhance the assessment of neuroimaging in the acute stroke setting to augment triage, rapid diagnosis, severity assessment, and outcome prediction. Lee et al⁸ compared the performance of 3 ML algorithms (logistic regression, support vector machine, and random forest) using magnetic resonance imaging (MRI) features with human ratings of diffusion-weighted imaging fluid-attenuated inversion recovery mismatch, which is prone to deficiencies in intrarater and interrater reliability. The ML-based methods had superior sensitivity (75.8% for logistic regression, 72.7% for support vector machine, and 75.8% for random forest assessment vs 48.5% for human assessment) and comparable specificity (82.6% for logistic regression, 82.6% for support vector machine, and 82.6% for random forest assessment vs 91.3% for human assessment) for identifying eligible individuals within the time window for acute thrombolysis (eg, within 4.5 hours of stroke onset). Rapid ASPECTS (iSchemaView), an ML automatic software tool based on the Alberta Stroke Program Early Computed Tomography (CT) Score (ASPECTS), demonstrated a higher level of accuracy for identifying early evidence of brain ischemia relative to experienced radiologists when using diffusion-weighted imaging as the standard.⁹ Furthermore, previous studies have shown that among individuals with subarachnoid hemorrhage, the combined use of electronic health record data and ML algorithms significantly outperformed standard models, with an overall improvement of 36% in the area under the curve for delayed cerebral ischemia detection and 9% and 18% improvement for discharge and 3-month outcome prediction, respectively.^10,11 The Ischemic Stroke Lesion Segmentation (ISLES) challenge, created in 2015 with the aim of encouraging globally diverse teams to develop ML-based tools for stroke evaluation, found a significantly better accuracy for ischemic core and penumbra identification by exclusively using deep learning methods when compared to the traditional threshold-based approach.¹²

AI tools have been used to identify large vessel occlusion, which is responsible for up to 40% of acute ischemic strokes, from CT angiography. The Automated Large Arterial Occlusion Detection in Stroke Imaging (ALADIN) study found that the Viz.ai algorithm v3.04, a convolutional neural network programmed algorithm used to detect occlusions at the M1 segment of the middle cerebral artery or internal carotid artery, was highly accurate and efficient.¹³ In addition to providing rapid and accurate detection of large vessel occlusion, which ultimately results in improved outcomes, the implementation of these AI tools has proven to be cost-effective.¹⁴

ML algorithms are being developed to provide automated assessment of hematoma expansion and perihematomal edema in people with intracerebral hemorrhage (Figure 2).¹⁵ Although their performance needs further evaluation, these new models have the potential to provide valuable information for risk stratification and clinical decision-making.

Remote Stroke Management

Despite advances in stroke management and improvements in outcomes in recent years, the centralization of acute stroke care facilities has led to an increase in rural–urban disparities in stroke outcomes in the United States. Although the causes of disparities in stroke care are multifactorial, insufficient access to advanced neuroimaging and neurointerventional resources and care delays are key factors contributing to the increased mortality rate among people in rural areas (18.6% vs 16.9% for individuals in rural vs urban settings, respectively).¹⁶ The shortage of specialized stroke clinicians in rural areas emphasizes a need for the development of novel approaches that can help improve patient management and outcomes. The use of telestroke, a branch of telemedicine, in rural settings is an innovative measure that has led to significant improvements in access to care, with equivalent stroke diagnostic accuracy when compared to bedside evaluations (radiographic evidence of stroke 78.3% at spoke vs 66.7% at hub; adjusted P=0.338).¹⁷ Several challenges remain, including a lack of community recognition of stroke symptoms, large distances to primary stroke centers, insufficient interhospital communication and coordination, and increased physician time needed for each assessment.⁵ The incorporation of geospatial AI tools could offer a unique opportunity to bridge these gaps.¹⁸

Stroke Prevention

Opportunities exist for AI use in primary and secondary stroke prevention. AI integration has been studied in electrocardiogram analysis to identify atrial fibrillation¹⁹ and to identify individuals at high risk in terms of CHADS2-VASC scores and AI-interpreted electrocardiography findings for further monitoring.²⁰ Screening the general population to optimize stroke prevention using clinical and laboratory data or biomarkers such as electrocardiographic abnormalities requires appropriate consideration to avoid overdiagnosis, overtreatment, and anxiety. However, if used appropriately, ML algorithms that analyze individualized data for risk assessment could have broad implications for both primary and secondary stroke prevention.

Stroke Recovery and Prognosis

AI presents new opportunities in poststroke rehabilitation. AI is being integrated into rehabilitation robotics, such as hand and leg exoskeletons, biofeedback therapies, virtual reality rehabilitation programs, and central–peripheral stimulation circuits.²¹ A recent randomized controlled trial studied a hand exoskeleton robot that incorporates AI detection of upper extremity electromyography (EMG) patterns to improve determination of movement intention and subsequent exoskeleton activation. People who trained with this robot showed improvements in upper extremity function and spasticity.²²

An important advantage of AI-driven models for poststroke prognostication is their ability to incorporate a large number of variables into their calculations and determine which variables are the most valuable for such predictions. In one such study, ML algorithms outperformed standard logistic regression models in predicting 3-year mortality rate after severe stroke.²³ Another study found that deep learning algorithms were superior to the Acute Stroke Registry and Analysis of Lausanne (ASTRAL) score in prediction of poor 90-day outcomes.²⁴ Given the capacity of these algorithms to elicit key variables contributing to stroke outcomes, AI-driven prognostic tools are anticipated to be integrated increasingly into the electronic medical record.

Future Directions

AI has great potential to improve medical care, but multiple technical, ethical, and legal questions remain unanswered. The lack of well-defined AI regulatory frameworks and trust among physicians has led to a reluctance to incorporate AI systems. This barrier must be addressed for AI progress to continue. Given the current rate of innovation in AI, proactive regulations to monitor and deploy these algorithms are needed to provide individuals, clinicians, and health care systems the confidence to adopt them. Another concern is the potential for algorithmic bias resulting in a lack of generalizability of these systems. Because most of the AI safety and efficacy studies are retrospective, prospective clinical trials with large participant samples evaluating and validating the use of these tools in stroke management should be a future priority. The incorporation of AI in stroke care should not be perceived as a replacement of the human workforce, but rather as a decision support tool to alleviate current shortcomings in care.

Limitations hindering the acceptance of current AI methods into routine clinical practice include the tradeoff between performance and explainability (eg, deep learning models have the best performance but are the least explainable), difficulty in assigning accountability for diagnostic errors, and vulnerability to malicious attacks. The importance of addressing these issues is critical, particularly in the medical field, where there is a demand for technologies that not only perform well but are also transparent, trustworthy, and explainable.

In the midst of a technologic revolution that is pushing boundaries and transforming health care, we must understand the strengths and limitations of AI and its branches to maximize their benefits. Particular focus should be placed on the development of human–AI collaborative approaches, which outperform either approach alone. This will be the key to redefining the current state-of-the-art diagnostic tools and treatment, and ultimately improving stroke outcomes.