INTRODUCTION
Artificial intelligence (AI) is revolutionizing healthcare, especially in stroke medicine where quick diagnosis and treatment are vital. The integration of AI technologies across the stroke care continuum, from initial diagnosis to outcome prediction, represents a significant advancement in our ability to deliver precise, personalized care in a timely manner. Recent years have witnessed an exponential increase in publications exploring AI applications in stroke medicine, with studies ranging from automated imaging analysis to prediction models for patient outcomes. The increase in research activity, especially since 2019, highlights the advancement of AI technologies and their acknowledged potential to address significant challenges in stroke care.
This scoping review provides an overview of the current research landscape on AI applications in stroke care, categorizing studies to assess the extent of existing advancements. It aims to summarize the available literature, offering insights into how AI is being utilized across various aspects of stroke management, including outcome prediction, stroke risk prediction, stroke diagnosis, etiology prediction, and complication or comorbidity prediction. With a systematic search using a novel automated methodology, articles were manually selected by the author with the aim to prioritize exploring more challenging areas that require additional investigation and development. The goal is to provide a clear understanding of the current state of AI in stroke care and its potential for advancing precision medicine.
Artificial Intelligence Models
Variety of algorithms has been used for AI. Support vector machines are trained for classification tasks by finding an optimal hyperplane that maximizes the margin between different classes in high-dimensional space [1]. Decision trees use a hierarchical structure where data is recursively split based on feature thresholds, creating an interpretable model that can handle both classification and regression tasks [2]. Random forests are an ensemble of these decision trees where each tree is trained on a random subset of the data and features [3]. Gradient boosting machines enhance predictive performance by sequentially training decision trees, where each tree corrects the errors of the previous one using gradient-based optimization [4]. Artificial neuron-based algorithms consist of interconnected artificial neurons that simulate biological neural networks by processing multiple weighted inputs and applying an activation function to produce an output [5]. Convolutional neural networks consist of multiple layers of these artificial neurons, including convolutional layers that extract spatial features, pooling layers that reduce dimensionality, and fully connected layers that perform classification [6]. These models are frequently used for image analysis. Other deep neural networks, such as recurrent neural networks and transformer-based models, are used for sequential data processing, particularly in natural language tasks [7,8]. Natural language processing models, such as Bidirectional Encoder Representations from Transformers and Generative Pre-trained Transformer, use self-attention mechanisms to understand context in text, allowing for advanced language comprehension and generation [9,10].
METHODS
Literature Search and Screening
A systemic literature search was performed on December 10, 2024, using National Library of Medicine database (PubMed) with the search term ((artificial intelligence [Title] OR machine learning [Title] OR deep learning [Title]) AND (stroke [Title] OR thrombectomy [Title] OR thrombolysis [Title] OR cerebral infarction [Title])). The search was performed with a custom Python script that automatically performs the pre-specified search using the National Institutes of Health (NIH) application programming interface and saving the results into a spreadsheet file. The script is available in the Supplementary Material 1. Search results were screened by assessment of their titles and abstracts. All scientific literature on the topic of using AI for ischemic stroke were included. Only original research articles were included for analysis, while other types of articles such as reviews, editorials, research protocols, or literature describing experience with AI software were excluded. Research mainly on the focus of rehabilitation after stroke was excluded. Studies only concerning hemorrhagic stroke were excluded. Non-human research was excluded. Manual selection was performed for final inclusion of this review.
Literature Classification
Search results were classified into the following categories after reading the title and the abstract of the literature: 1) outcome prediction, 2) stroke risk prediction, 3) stroke diagnosis, 4) etiology prediction, 5) complication or comorbidity prediction, and 6) others. The outcome prediction category included studies that used AI to predict outcome for patients with stroke. Stroke risk prediction included research that aimed to develop models for occurrence of stroke. Studies classified as stroke diagnosis included those which used imaging, clinical variables, or novel modalities to aid in the diagnosis of stroke. Research that used AI to classify or predict the etiology of stroke was included in the etiology prediction category. The complication or comorbidity prediction category included studies that predicted the occurrence or presence of other diseases. Articles that did not fit into any of these categories were placed in the ‘other’ group.
Initial Literature Analysis Using Artificial Intelligence
Large language models were used for initial screening and categorization of search results. The results provided by the models were reviewed and edited by the corresponding author. Two models were used sequentially for each row of the spreadsheet file that contained the search results. The first model assessed the abstract to see if the article meets any of the exclusion criteria. If the article was not excluded, the article was sent to the second model. The second model categorized the article into the 6 pre-specified categories. The results were saved as new columns and saved as a separate spreadsheet file. The script is available in the Supplementary Material 2.
Literature Search Results
A total of 821 results were retrieved from the search. Of these results, 505 articles (61.5%) were included for analysis after exclusion (Fig. 1). Outcome prediction was the most studied category (198 articles, 39.2%), followed by stroke diagnosis (166 articles, 32.9%), stroke risk prediction (92 articles, 18.2%), etiology prediction (20 articles, 4.0%), and complication and comorbidity prediction (12 articles, 2.4%). On further classification of articles inside the outcome prediction category, functional outcome was the most studied topic (101 articles, 51.0%), followed by mortality (26 articles, 13.1%), and stroke recurrence (16 articles, 8.1%). The number of articles on AI and stroke has grown rapidly since 2019 (Fig. 2). However, the rate of increase has plateaued since 2022. Outcome prediction studies have consistently been published from 2014, while stroke diagnosis has also been studied extensively, especially in 2022. A total of 52 articles were manually selected for this review (Supplementary Table 1).
STROKE OUTCOME PREDICTION
Functional Outcome
One of the earliest large-scale studies on the topic of functional outcome prediction used the Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands Registry to retrospectively develop models to predict good reperfusion (modified Thrombolysis in Cerebral Infarction≥2b) and functional independence (modified Rankin Scale≤2) at 3 months using baseline variables with or without treatment variables [11]. Random Forests, Support Vector Machine, Neural Network, and Super Learner were used and compared with logistic regression models. The area under the receiver operating characteristic curve (AUC) was 0.53–0.57 for reperfusion prediction and 0.77–0.79 for functional independence prediction. Upon adding treatment variables, the AUC increased to 0.88–0.91 on functional independence prediction, indicating the importance of reperfusion treatment in functional outcome. Another earlier large-scale study included 2,604 patients to develop models to predict functional independence at 3 months [12]. Deep neural network, Random Forest, and logistic regression was used to build models and compare outcomes with the Acute Stroke Registry and Analysis of Lausanne (ASTRAL) score. Baseline clinical variables were used as input. The deep neural network model showed an AUC of 0.888, which was significantly higher compared to the ASTRAL score (0.839, P=0.001). More recently, a study has shown that models incorporating multimodal inputs have shown higher performance compared to single-modality models [13]. An ensemble model including 3D convolutional neural networks of diffusion-weighted imaging and fluid-attenuated inversion recovery imaging performed significantly better (AUC of 0.830) than models that used either a diffusion coefficient map, fluid-attenuated inversion recovery imaging, or clinical data as predictors. When compared to stroke neurologists, AI models performed similarly when only clinical variables were used and may perform better when imaging is also used [14,15].
Reperfusion Therapy Outcomes
AI was used to predict outcome measures associated with acute reperfusion therapies. Successful recanalization after the first thrombectomy effect, also known as first-pass effect, was predicted using pre-treatment CT and MR imaging [16]. Using a cohort of 326 patients, a hybrid transformer model with non-local and cross-attention modules was trained and validated. No manual segmentation was needed to use the model. The MR-based model showed an AUC of 0.797 and the CT-based model showed an AUC of 0.805. Models using only clinical variables were developed to also predict the first-pass effect in another study [17]. Random Forest, Support Vector Machine, and logistic regression were applied, yielding final accuracies between 65.8% and 67.1%. Of note, feature attribution analysis revealed that the most influential variable was the interventionist’s choice of an aspiration catheter as the thrombectomy device, underscoring the potential bias introduced by human judgment in the model’s predictions. Success after intravenous thrombolysis was predicted with clinical variables and AI [18]. The AUC was 0.765, with the initial NIH Stroke Scale (NIHSS) being the most important variable in outcome prediction, followed by glucose, neutrophil, white blood cell, and blood urea nitrogen levels.
Stroke Recurrence
Stroke recurrence was predicted using AI models. Using a large public health care-based dataset (41,975 patients with stroke), models were developed to predict early (≤90 days), late (91–365 days), and long-term (>365 days) recurrence risk [19]. Of the included patients, 16% (5,932 of 36,114) suffered stroke recurrence during follow-up (median 2.69 years). The AUCs were 0.76 for early, 0.60 for late, and 0.71 for long-term recurrence risk prediction models. The most important predictors were time from previous stroke, Barthel Index, atrial fibrillation, dyslipidemia, age, diabetes, and sex. AI models were significantly better at prediction than the Cox model. Another study that developed recurrence prediction models focused on patients with symptomatic intracranial cerebral artery stenosis [20]. The AUCs ranged from 0.813 to 0.912 for 5 models on 10-fold cross-validation. Feature attribution analysis showed radiologic characteristics of high-resolution vessel-wall imaging were important in recurrence.
STROKE RISK PREDICTION
Initial and Recurrent Stroke Risk
Prediction modelling of occurrence of stroke has been studied extensively with AI algorithms. A study investigated stroke risk in a large prospective cohort of 3,435,224 U.S. patients with multimorbidity, utilizing traditional clinical risk scores, a clinical multimorbid index, and AI approaches [21]. While common clinical scores (e.g., CHADS2, CHA2DS2-VASc) showed moderate predictive ability, the clinical multimorbid index and AI-based algorithms demonstrated superior discriminative performance (c-index of 0.866). The AI models effectively accounted for the dynamic nature of multimorbidity and showed better calibration and decision curve analysis compared to clinical rules. In addition to predicting initial stroke risk, efforts have also been made to predict the risk of recurrent strokes using AI [22]. A study developed prediction models based solely on patient-reported data easily obtained at home, eliminating the need for costly clinical or imaging data. The best model achieved an AUC of 0.70, with good probability calibration. Key predictive factors included family and housing circumstances, rehabilitative measures, age, diet, blood pressure levels, and the frequency of family doctor home visits.
Stroke Risk Prediction in Cardiac Patients
Patients with cardiac diseases are at higher risk of developing stroke. Research has been conducted to predict stroke occurrence in patients with atrial fibrillation, coronary artery disease, and those who have undergone percutaneous coronary intervention. One of the earliest published studies on risk prediction in patients with atrial fibrillation developed a 2-year prediction model with data of 1,864 patients and achieved an AUC of 0.71–0.74 [23]. Another study leveraged continuous remote monitoring data from cardiac implantable electronic devices to evaluate signatures of atrial fibrillation burden for near-term stroke risk stratification with 3,185 patients [24]. Using convolutional neural networks, random forests, and LASSO logistic regression, the study found that these models outperformed the conventional CHA2DS2- VASc score with the highest AUC of 0.702. In a case-control study involving 8,624 patients, machine learning (ML) models for predicting coronary heart disease and ischemic stroke achieved AUCs of 0.905 and 0.889, respectively, identifying key predictors such as age, brachial-ankle pulse wave velocity, hypertension, and low-density lipoprotein cholesterol [25]. Another study analyzing data from 17,356 patients developed random forest models to predict ischemic stroke after percutaneous coronary intervention, with an AUC that exceeded logistic regression models at all time points, reaching 0.662 at 6 months and 0.702 at 5 years [26]. Additionally, a cohort of 5,757 coronary artery disease patients receiving coronary revascularization showed that a CatBoost model outperformed logistic regression for predicting postoperative stroke, achieving an AUC of 0.760 in the testing set, with the Charlson Comorbidity Index identified as the most important predictor [27].
Carotid Plaque Imaging for Stroke Risk Assessment
Carotid plaque imaging and advanced analytics have demonstrated utility in stroke risk prediction. Automated segmentation of grayscale carotid plaques achieved high classification accuracies of 95.08% and 93.47% for far and near walls, respectively [28]. Deep learning models applied to color Doppler images classified carotid plaques as high-risk or stable with a maximum accuracy of 93.81%, using pretrained Inception V3 and VGG-16 models fine-tuned for this task [29]. MRI-based analyses showed 94.81% accuracy in identifying carotid plaques with YOLOv3 models, emphasizing operator independence and the superior soft-tissue contrast of MRI in stroke risk assessment [30]. Plaque morphology shown in CT was also utilized [31]. A combined model incorporating clinical factors such as age, body mass index, and a history of transient ischemic attack or stroke with plaque morphology variables like perivascular adipose tissue and lipid-rich necrotic cores achieved an AUC of 0.759 for 30-day stroke or mortality.
Fundus Imaging for Stroke Risk Assessment
Fundus imaging has been utilized with AI to predict stroke occurrence. A study involving 250 participants with atrial fibrillation analyzed fundus images captured at wavelengths of 548, 605, and 810 nm using 3 classical deep neural networks (Inception V3, ResNet50, SE50) [32]. The models achieved an accuracy of over 78%, with the 605 nm wavelength showing the most stable detection performance in prediction of stroke occurrence. Multi-spectral combinations of images yielded higher AUC scores compared to single-spectrum models. Another study analyzed retinal vasculometry data from 88,052 and 7,411 participants across 2 large cohorts to assess its potential in improving risk algorithms for incident stroke, myocardial infarction, and circulatory mortality [33]. The addition of retinal vasculometry data to the Framingham risk scores did not significantly enhance prediction; however, a simpler model using retinal vasculometry data, age, smoking status, and medical history performed equally or better than Framing risk scores, with C-statistics ranging between 0.75–0.77 for circulatory mortality prediction. Furthermore, a deep learning-based system trained on retinal biomarkers from fundus images achieved AUC of 0.83 and 0.93 in 2 datasets, demonstrating superior performance over traditional benchmarks in predicting stroke risk and timing of occurrence [34].
STROKE DIAGNOSIS
Large Vessel Occlusion Detection
AI was used to automatically estimate Alberta Stroke Program Early CT score [35]. Numerous commercially approved software solutions offer this functionality and are frequently used in clinical settings [36]. Studies have shown that these software improved efficiency in imaging analysis [37,38]. Similarly, AI models were developed and validated for detection and segmentation of ischemic lesions in non-contrast brain CT imaging [39,40]. Models focused on large vessel occlusion detection have also been extensively studied and validated, currently available as commercial software certified for use in clinical settings [41-43]. Performance of large vessel occlusion detection using AI was studied for CT imaging obtained from mobile stroke units [44]. Automatically detecting large vessel occlusion at the mobile stroke unit before arrival at the emergency room could facilitate faster reperfusion times, leading to better treatment outcomes.
Ischemic Lesion Segmentation
Ischemic lesion segmentation on non-contrast brain CT was performed with AI [39,45,46]. One of the main challenges in developing a segmentation model for non-contrast brain CT is the low correlation of ‘ground truth’ segmentation between experts. The decision of the ‘right’ lesion after manual delineation by experts was performed by reaching a consensus [46]. Although with a limited number of images (260 CT studies), developing a model with a dataset consisted of randomly choosing 1 out of 3 experts’ segmentation for each CT study showed better performance than a merged segmentation map [39]. An alternative approach, despite its limitations, is indirectly deriving the segmentation map using expert-labeled ischemic lesion volumes from diffusion-weighted MRI scans [45].
AI-Enhanced Imaging
AI-accelerated ultrafast MRI, which takes about a fourth the time to acquire compared to traditional MRI, was validated [47]. Interchangeability on detection of cerebral infarction of the ultrafast MRI compared to traditional MRI was assessed with 3 radiologists as primary outcome. Secondary findings such as microbleeds or neoplasms, vascular territory of the infarction, and image quality assessment were also compared. There was no significant difference in the assessment results by the radiologist between the 2 modalities. A CycleGAN-based deep learning model was developed to generate high-resolution synthetic time of flight (TOF) images from time-resolved MRA for improved collateral evaluation in acute ischemic stroke [48]. Image quality assessments showed that synthetic TOF images had better overall quality compared to time-resolved MRA but were inferior to real TOF in some areas. Clinical validation demonstrated synthetic TOF enhanced diagnostic confidence and reduced decision time in detecting large-vessel occlusion, showing the potential for use in stroke centers.
Other Studies on Stroke Detection
A study was performed to automatically recognize stroke by using AI on transcription of medical helpline calls [49]. The model performed significantly better than the call-takers with a sensitivity of 63%. The most important words in identification of stroke were ‘ambulance,’ ‘blood clot,’ ‘left,’ ‘right,’ and ‘double vision.’ AI was used in the analysis of movement data of patients to diagnose stroke. The pronator drift test, an important neurologic examination for stroke, was performed using accelerometer signals from wearable devices on the wrist [50]. Stroke was predicted using ML of signals acquired from the sensors. The AUCs ranged from 0.913 to 0.975 in the prediction of stroke. In another study, based on the Face Arm Speech Test clinical test, multi-modal deep learning models were developed to analyze the movement and speech data of the patients under specific instructions (i.e., raise both arms or pronounce a pre-specified sentence) [51].
ETIOLOGY PREDICTION
Determining the etiology of stroke is crucial for effective secondary prevention. However, despite thorough evaluations, the cause remains unknown in over a third of stroke patients. Innovative approaches leveraging AI could help address this unmet need.
Detection of Atrial Fibrillation
Using demographics, comorbidities, vitals, laboratory results, and echocardiograms, a model was developed to classify between cardioembolic and non-cardioembolic strokes [52]. The model was applied to patients with embolic stroke of undetermined source and compared the results with eventual diagnosis of atrial fibrillation in these patients. The model’s prediction was associated with eventual atrial fibrillation detection. Electrocardiography during sinus rhythm was used for its potential in prediction of hidden paroxysmal atrial fibrillation [53]. A model was developed for prediction of underlying atrial fibrillation with electrocardiography results from patients with paroxysmal atrial fibrillation. This model was applied to patients initially classified as embolic stroke of undetermined source and compared with prolonged ambulatory monitoring. The probability of atrial fibrillation in the prediction model was significantly associated with detection of atrial fibrillation on prolonged monitoring.
Imaging-Based Subtype Classification
Automatic subtype classification was performed with AI using diffusion weighted MR imaging and presence of atrial fibrillation, achieving performance comparable to expert consensus [54]. By utilizing the presence of atrial fibrillation and patterns of infarction in diffusion weighted imaging, the AI model showed high agreement (72.9% agreement) with the expert consensus. A model using only the diffusion weighted imaging data showed a lower agreement (58.1%) with the expert consensus. These agreement values of the models were comparable to the agreement between the experts, which was only 76%.
Thrombi-Based Subtype Classification
AI was used for analysis for thrombi to predict etiology of stroke. In a multicenter study, whole-slide scanned images of immunohistochemically-stained thrombi retrieved from mechanical thrombectomy were used for detection of underlying cancer as the etiology for stroke [55]. The model utilizing the platelet-stained slides showed a high AUC of 0.949 on external validation, reconfirming the importance of platelet in cancer-associated thrombi in stroke. Ex-vivo multiparametric MR imaging was also studied for prediction of dichotomized red blood cell component of the thrombi [56]. In this experimental study, MR imaging was obtained with the thrombi retrieved from thrombectomy. After the MR imaging, the thrombi were stained to analyze the red blood cell component. A model was developed using the MR image to predict if the thrombi contain a high percentage of red blood cells. The AUC was 0.84 on cross-validation.
COMPLICATION OR COMORBIDITY PREDICTION
Stroke-Associated Pneumonia
The most studied complication after stroke was pneumonia. An XGBoost model was developed to predict stroke-associated pneumonia in 3,160 patients with stroke, achieving an AUC of 0.841 and outperforming traditional risk scores [57]. This model relied on 6 commonly available clinical variables: age, pre-morbid modified Rankin Scale score, NIHSS at admission, fasting blood glucose level, sex, and history of atrial fibrillation. Occurrence of pneumonia within 1 year after discharge was predicted with a large dataset of 5,754 hospitalized stroke patients [58]. The C-index was 0.787 on the test set, and the most important features were Glasgow Coma Scale score, age, and length of hospital stay.
Acute Kidney Injury
AI was used to predict acute kidney injury, analyzing data from 3,920 patients [59]. Among 9 model, the XGBoost model demonstrated superior performance, achieving AUCs of 0.940 and 0.887 in internal and external validation sets, respectively. SHapley Additive exPlanations analysis identified glomerular filtration rate, low-density lipoprotein, total cholesterol, hemiplegia, and serum potassium as the most important predictors.
Hidden Coronary Artery Disease
A study developed and validated ML models to predict hidden coronary artery disease in patients with stroke using clinical variables [60]. A total of 2,058 patients were included for this study, and the best performing model showed an AUC of 0.763 on validation. Additionally, major adverse cardiovascular events were more frequent for the patients who were predicted to have a high possibility of coronary artery disease during 5-year follow-up. This study highlighted the potential of AI to identify hidden coronary artery disease and inform risk stratification in stroke patients.
OTHER STUDIES
Two recent studies highlight the potential of AI to optimize acute stroke care pathways. The first study developed a mobile app integrating ML-based prediction models with historical data on air and land medical transport to guide healthcare providers in Northwestern Ontario on the most efficient patient transfer options for stroke care [61]. Using a decision tree model trained on over 70,000 transport records, the app accurately accounts for variables like patient location, available treatments, and imaging facilities, aiming to improve outcomes and reduce healthcare costs. However, the study was based on retrospective data, and the usage results of the developed app is needed. The second study explored the feasibility of Wi-Fi fingerprinting with ML algorithms, such as random forest and support vector machines, to create a real-time location system for streamlining acute stroke endovascular interventions [62]. With an accuracy of 98% in tracking patient and staff movements across hospital zones, this technology could significantly enhance workflow efficiency during time-sensitive procedures like mechanical thrombectomy.
AI may have potential in optimizing mechanical thrombectomy for acute stroke patients by addressing procedural challenges. A model demonstrated that acute internal carotid artery angles (≤90°) are associated with longer mechanical thrombectomy durations, achieving high segmentation accuracy (Dice scores of 0.94 for the aorta and 0.86 for internal carotid artery) and robust performance in predicting internal carotid artery angles (AUC of 0.92). Similarly, a model trained on anatomical features extracted from CT angiography effectively predicted difficult transfemoral access during mechanical thrombectomy, outperforming expert assessments with an AUC of 0.76.
CONCLUSION
While AI shows significant promise, its role in healthcare, including stroke management, is still evolving. Continued focus on methodical validation, ethical considerations, and practical implementation will determine the extent to which AI can be effectively applied to improve clinical outcomes. Although further evidence is needed, AI has the potential to play an important role in enhancing stroke care and addressing current challenges in the field.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement
Print
