Artificial Intelligence in Neurology - Cerebral Ischemia

Time is brain! Artificial Intelligence implementation in cerebral ischemia decrease care latency improving outcomes!


  • Stroke has a significant burden on cost and quality of life
    • Cost: Global Cost is almost $861 Billion which is around 1.21% of global GDP.
    • Almost 89% of global stroke death and disability combined resides in low income countries.
    • Quality of Life: Most common cause of Disability
      • Cost 200000$ per patient per quality of life gained (ref)
  • AI/ML has the potential to reduce cost by:
    • Early detection
    • Early intervention
    • Intelligent automation
    • Intelligent Rehabilitation
  • AI/ML has already been recognized by FDA for De Novo Clearance
  • AI/ML in Stroke is the first to be reimbursed by CMS via NTAP (New Technology Add-on Payment (NTAP)
  • Additional, Stroke has a system of care approach
    • AI/ML can improve care coordination
    • AI/ML can help in triaging patients and selecting patients for intervention
    • With a virtual environment, AI/ML can create a real-time environment for EMS providers.
  • Review the introduction - before continuing to read

AI/ML in Neurovascular disorder

  • AI/ML is deployed within the continuum of care of neurovascular disorders
  • As we need to identify each step in the process and use automation and intelligence
  • The first step in Focal Neurological deficit is to detect cerebral hemorrhage
  • This is discussed in detail in the chapter

AI/ML in Cerebral Ischemia

  • Machine Learning is the part of artificial intelligence that collects data, automates the process, and modifies it based on previous experiences.
  • AI/ML is deployed at various stages during the stroke care
  • We have divided the applications in the patient journey
  • Saving one minute for patient of ischemic stroke equals to
    • Onset to tPA: 1.85 days of extra healthy life
    • Onset to EVT: 4.2 days of extra healthy life
    • TICI 2b-3 revac: 4.3 days of extra healthy life
    • Patient <55 years old with NIH >10: >1 week of extra healthy life
  • Patient selection (Detection): The convolution neural network (CNN), a subtype of ANN has been used to classify and analyze imaging.
    • Rule out hemorrhage
    • ASPECTS score (tissue at risk)
    • Large vessel occlusion (automated cerebral large vessel detection)
    • Ischemic Penumbra (automated cerebral perfusion)
  • AI/ML algorithms can be deployed by
    • Training a new algorithm
    • Use a generic algorithm and modify it as per requirement
  • Hassan et. al., incorporated software and compared transfer time of patients from primary stroke center to comprehensive stroke center.
    • The usual time from arrival to cath. lab was 117 minutes.
    • After incorporation of AI software
      • Time of transfer was reduced by almost 22.5 minutes.
      • Total length of stay in hospital or neurological ICU was reduced. (ref)
Improved Patient selection for thrombolytics - (Not just endovascular intevention)
  • EXTEND numbers, NCT00887328 and NCT01580839.
    • A multicenter, randomized placebo-controlled trial showed that with the help of automated perfusion imaging the use of alteplase between 4.5 and 9 hours after stroke results in higher number of patients with no neurological deficit as compared to placebo. (ref)
  • ECASS4-EXTEND, and EPITHET are other trials to extend the window to use thrmbolytics. (ref)

Automated Diagnosis of Cerebral Ischemia

  • AI/ML-based algorithm can help in making a diagnosis especially
    • Hospitals where expert Neuroradiologist is not available
    • During night shifts
    • Hospitals with limited resources.
  • AI/ML can help in
    • Differentiating ischemia from hemorrhage
    • Quantification of infarct
    • Prediction of function loss
  • Listed in table 2 are some ML-based algorithms for diagnosing ischemic strokes.
TOAST classification divides acute ischemic stroke into five major types. 1. Atherosclerosis of large artery 2. Embolism from heart 3. Occlusion of small vessel 4. Stroke due to known cause 5. Stroke due to undetermined cause

Alberta Stroke CT Score (ASPECTS)

  • It is a ten points score based on non-contrast CT
    • Patients with middle cerebral artery stroke
    • One point is deducted with each region involved (Figure 1)
    • A score of less than 7 is associated with the worst prognosis
  • Adjusted for posterior circulation (pc-ASPECT)

Figure 1: ASPECT Score

notion image


  • Calculation of ASPECTS with AI/ML-based algorithms.
  • Gregory W Albers and his team compared rapid ASPECTS and CT ASPECTS of physicians with Diffusion-Weighted imaging (DWI).
    • Rapid ASPECTS has less margin of error
    • Rapid ASPECTS is better in diagnosing at an early stage of stroke.
  • Noke Guberina and his team compared automated ASPECTS with three clinicians.
    • Automated ASPECTS sensitivity was 83%
    • Automated ASPECTS enhance the plausibility of diagnosis
  • Some of the RAPID ASPECTS algorithms are listed in Table 2

Automated detection of large vessel occlusion (LVO)

  • AI/ML algorithms can help the stroke patient with LVO
    • Diagnosis assistance
    • Saving time
    • Analyzing eligibility for endovascular therapy
  • Algorithms available to detect large vessels occlusion are
    • Convolution neural network (CNN)
    • Random Forest Model (RF)
    • CNN model is superior to the RF model

Table 1: Automated Large Vessel Occlusion (LVO) detection

Shalini A. Amukotuwa, (ref)
CT scan of patients with ischemic stroke
Automated LVO detection (RAPID - CTA)
Two Neuroradiologists
RAPID -CTA had better sensitivity
Sven P R Luijten, (ref)
Data from the MR CLEAN Registry and PREST
Automated LVO detection
Assessment by an expert neuroradiologist
The algorithm performs better for proximal LVO but varies with location
Zhicai Chen et. al., (ref)
CT Scan of Stroke ischemic patient
Automated LVO detection by an artificial neural network (ANN) in the pre-hospital triage stage
Established pre-hospital prediction scale
The AUC and accuracy of the ANN algorithm were higher than the established prediction scale.


  • It is important to differentiate between reversible injury, ischemic injury, and infarct.
  • Automated quantification of ischemic penumbra.
  • Common methods of segmentation
    • Relative cerebral blood flow (rCBF)
    • Diffusion-weighted imaging
  • Maier and his team reported
    • RF and CNN models outperformed as compared to all other models
    • RF and CNN models underperformed as compared to the radiologist
  • Dice score is the widely used score to compare the efficacy of AI/ML algorithms.
    • A score is closer to one means higher efficacy
  • ATLAS (Automated Tree Learning Anomaly Segmentation)
    • ML algorithm based on decision tree
    • Used for anterior and middle cerebral artery circulation
    • The dice score was 0.6122 (108 patient’s data)
  • Table 2 lists some of the algorithms for automating the segmentation process.
The Decision Tree algorithm is a supervised algorithm that learns from simple decisions and infers from them.


  • Borderline sensitivity
  • High false-negative rate
  • Simple tasks are challenging such as
    • Artifacts detection
    • Accounting patients’ motion
    • Old lesion
  • Estimation of lesion size
    • CNN models overestimate for smaller lesions and vice versa

Table 2: AI/ML-based algorithms

Automated ASPECT analysis
Automated Stroke Diagnosis
Automated Segmentation of Infarction
Convolution Neural Network
Frontier ASPECTS Prototype
Random Forest
iSchema View (FDA approved)
Visual Geometry Group from Oxford (VGG-16),
Random Forest Classifier
Google Net
Generalized Linear Models
Residual Network with 50 layers (ResNet-50)
Random Decision Forests
Support Vector Machine
Sparse Autoencoder (SAE) layers
Support Vector Machine (SVM)


  • The aim of ischemic stroke management is to save the ischemic tissue as much as possible.
  • There are two modalities that are common in the management of ischemic stroke.
    • IV tissue plasminogen activator
    • Emergency endovascular techniques
  • AI/ML-based algorithm can help in
    • Selection of patient based on quantified ischemic penumbra.
    • Outcome prediction
    • Prediction of complication
    • Prediction of post-stroke quality of life.
  • AI/ML can help in triaging patients
    • Home-based 3D imaging technologies can detect facial asymmetry
    • The accuracy of video-based applications are up to 87%
  • Critical timelines for stroke management are
    • 4.5 hours for tPA
    • Up to 6 hours for endovascular thrombectomy
  • The delay in first aid provision can lead to permanent functional impairment.
  • In a pre-hospital context, AI/ML can help in reducing the time from event to first aid.
    • Using FAST-ED and eFACE assessments.
    • Automated traffic efficient routes availability for an ambulance.
    • Telestroke platforms
    • Automated alerts to the hospital stroke team
    • Automated registration of patient
    • Automated allocation of resources

Selection of Patient

  • It is important to select the treatment modality for the patient as early as possible.
  • Early treatment selection can help us to avoid unwanted complications.
  • Inclusion and exclusion of patients are based majorly on the volume of infarction.
  • The outcome prediction model can help in selecting treatment modalities for patient
  • ML algorithms can quantify the volume of an infarct with similar accuracy to DWI.
    • Support Vector Machine (SVM) (AUC of 74%)
    • 3-dimensional pseudo-continuous arterial spin labeling (pCASL) (92% accuracy)
    • SPOT is an algorithm based on a regression tree model
      • Helps in the selection of patients for endovascular therapy
      • Predicts 90 days modified Rankin Scale
      • AUC was 0.952 (36 patients’ data)
    • Maarten G Lansberg did a study on the automated selection of patients for reperfusion therapy (ref)
      • RAPID algorithm
      • EPITHET and DEFUSE trial data used
      • Patient profiles that can respond to tPA can be identified by using RAPID algorithms

Predicting Prognosis

  • AI/ML algorithm can help in predicting prognosis by analyzing
    • Clinical data
    • Imaging
    • Quantification of ischemic penumbra
    • Automated perfusion techniques
  • Predicting prognosis can help in
    • Improving survival
    • Devising patient-centered treatment plan
    • Patient-centered exercises during post-stroke rehabilitation

Predicting Hemorrhagic Conversion

  • Post thrombolysis/thrombectomy bleed is the most common complication.
  • AI/ML can help in predicting hemorrhagic conversion
    • Swarm algorithm
    • Partial least square(PLS) algorithm
    • Genetic algorithms.
  • Chihhsiong and team have developed a modified PSO algorithm based on PLS and genetic algorithm
    • Helps in predicting hemorrhage based on patient’s co-morbidities
    • Initial analysis showed an accuracy of 86%
  • Choi et. al., compared different algorithms’ accuracy in predicting hemorrhagic conversion
    • Binary Logistic Regression (BLR) - 83.7% accuracy
    • Support Vector Machine (SVM) - 85.6% accuracy
    • Extreme Gradient Boosting (XGB) - 83.4% accuracy
    • Artificial Neural Network crude model (ANN_crude) - 87.8% accuracy

Predicting Malignant Cerebral Edema

  • Malignant cerebral edema is a life-threatening complication.
  • Urgent hemi-craniectomy is required because of herniation risk
  • Foroushani et. al. did a pilot study
    • Fully Connected and Long short-term memory (LSTM) neural networks were trained
    • Data used: Clinical and imaging
    • Compared with regression models and EDEMA score.
    • The fully connected network didn’t perform well.
    • LSTM identified all the cases of malignant EDEMA by 24 hours after stroke (fewer false positives)

Predicting Tracheostomy

  • Post-stroke tracheostomy is a part of continued medical care due to airway compromise.
  • The identification of the tracheostomy need during the early period of medical care can help in protecting the airway preemptively.
  • Some of the algorithms available are
    • Random forest (AUC 0.74)
    • Gradient Boosting Machine (AUC 0.75)
    • XGBoost (AUC 0.74)

Predicting post stoke pneumonia

  • Pneumonia occurring within the first week is usually associated with stroke.
  • Post-stroke pneumonia is one of the common complications that can increase the length of the stay.
  • Some of the algorithms that can help in predicting post-stroke pneumonia
    • Multiple layer perception (MLP)
    • Recurrent neural networks (RNN)
    • Logistic regression,
    • Support vector machines
    • Extreme gradient boosting
  • The Attention-augmented gated recurrent unit (GRU) based on RNN performed better both at 7 and 14 days.

Predicting post-stroke cognitive performance

  • Prediction of post-stroke cognitive function can help in
    • Patient-centered treatments
    • Goal-oriented neuro-rehabilitation exercise
    • Targeted cognitive function
    • Helps in predicting patient’s quality of life
  • Chauhan and her team compared four models
    • Assessed Accuracy and effect of sample size data redundancy.
    • The hybrid model performed better than others. (Table 3)

Table 3: Algorithms to predict post-stroke cognitive function

Area Under Curve
Prediction of long-term language deficit (R-squared)
Ridge Regression Method (RR) with PCA (principal component analysis)
Support Vector Regression (SVR)
Convolution Neural Network (CNN)
Hybrid Model (RR plus CNN)
Extracted from: Chauhan S, Vig L, De Filippo De Grazia M, Corbetta M, Ahmad S and Zorzi M (2019) A Comparison of Shallow and Deep Learning Methods for Predicting Cognitive Performance of Stroke Patients From MRI Lesion Images. Front. Neuroinform. 13:53.

Artificial Intelligence Assisted Neuro-rehabilitation

  • The global neuro-rehabilitation devices market can reach up to $2.1 billion by 2026 (ref)
  • Tele-rehabilitation using socially assistive robots can aid in the emotional support and engagement of patients.
  • Two models of remote rehabilitation
    • The model that is Home centered
    • The model that reduces the time of patients in institutions
  • Limitations of existing neuro-rehabilitation model of care
    • Rehabilitation tools are not adaptive
    • Loss of patient’s engagement and interest
    • Costly
    • Dependent on human resources
    • Required clinical settings
  • mHealth apps and IoT devices can help in resolving these problems.
    • Adjustable as per patient needs
    • Increase difficulty in exercises gradually based on results
    • Monitor progress
    • Communicate with healthcare providers
    • Provides real-time quantifiable data
    • Patient-centered exercise protocol
  • Some of the mHealth apps and IoT devices are listed in Table 4.
  • In resource-limited countries, connected rehab centers can be introduced
    • Limited supervision
    • Fewer resources will be required
    • Cost-effective


  • Most AI/ML-based devices still need human supervision
  • Provision of technology as per patient’s needs a certain framework
  • Regulatory hurdles
  • Data privacy and security

Table 4: Artificial Intelligence-based mHealth applications and health IoT in Neuro-rehabilitation

4A: mHealth Applications in Neuro-rehabilitation

mHealth Apps
Helps patients to maintain balance. Patients exercise by moving their feet around the clock face.
A game-based app to help patients. Improves speech, finger dexterity, and also the mood of the patient.
Helps with speech and language. Reads texts aloud and helps patients with dyslexia.
A physiotherapy app helps patients to do exercise along with reminders about their appointments.
Videos depicting the normal swallowing process. Clinicians can use it to educate people.
Stroke-specific app showing videos about limb rehabilitation. Heps clinicians refine reasoning and rehab techniques.
Helps pro athletes refine their technique. Clinicians can ask their patients to make videos to watch later on.
Extracted from: Faheem, F; Kalia, J, mHealth in Neurology - An Introduction. Neurology Pocketbook. Published January 31, 2022. Accessed February 4, 2022. ‌

4B: Health IoT devices in Neuro-rehabilitation

Table 5: Commercially Available Software Applications For Artificial Intelligence In Stroke

Regulatory Approval
Imaging Modality
Automated ICH Detection
Automated ASPECTS
Automated LVO Detection
Automated Cerebral Perfusion
Cerebral Aneurysm
Care Coordination
FDA Approved
CT and CTA
Large Vessels Occlusion and ICH detection Rapid Triage and Communication solutions
FDA Approved NTAP Reimbursed
Large Vessels Occlusion and ICH detection Communication and Co-ordination solutions
FDA Approved NTAP Reimbursed
Pulmonary embolism
FDA Approved
Images analysis and quantification of disease icompanion - mhealth app for remote monitoring
FDA Approved
Segmentation and lesion’s volume quantification on FLAIR MRI
FDA Marked
Diagnostic solution
CE Marked
Diagnostic and treatment support.
Diagnostic and care communication support
Large vessels occlusion detection
Diagnostic solution for stroke
Segmentation and lesion’s volume quantification Perfusion mapping solutions
Automated vessel segmentation and brain perfusion mapping
Automated processing and visualization Provides perfusion and diffusion data
FDA Marked
Diagnostic and workflow support Quantify and monitor brain tumor volume
Rapid triage solution

Further Reading

  • Shafaat O, Bernstock JD, Shafaat A, et al. Leveraging Artificial Intelligence in Ischemic Stroke Imaging. Journal of Neuroradiology. Published online May 2021. doi:10.1016/j.neurad.2021.05.001
  • Kamal H, Lopez V, Sheth SA. Machine Learning in Acute Ischemic Stroke Neuroimaging. Frontiers in Neurology. 2018;9. doi:10.3389/fneur.2018.00945


  • Meretoja A, Keshtkaran M, Tatlisumak T, Donnan GA, Churilov L. Endovascular therapy for ischemic stroke. Neurology. 2017;88(22):2123-2127. doi:10.1212/wnl.0000000000003981
  • Pop N, Tit D, Diaconu C, et al. The Alberta Stroke Program Early CT score (ASPECTS): A predictor of mortality in acute ischemic stroke. Experimental and Therapeutic Medicine. 2021;22(6). doi:10.3892/etm.2021.10805
  • Albers GW, Wald MJ, Mlynash M, et al. Automated Calculation of Alberta Stroke Program Early CT Score. Stroke. 2019;50(11):3277-3279. doi:10.1161/strokeaha.119.026430
  • Guberina N, Dietrich U, Radbruch A, et al. Detection of early infarction signs with machine learning-based diagnosis by means of the Alberta Stroke Program Early CT score (ASPECTS) in the clinical routine. Neuroradiology. 2018;60(9):889-901. doi:10.1007/s00234-018-2066-5
  • Shafaat O, Bernstock JD, Shafaat A, et al. Leveraging Artificial Intelligence in Ischemic Stroke Imaging. Journal of Neuroradiology. Published online May 2021. doi:10.1016/j.neurad.2021.05.001
  • Monteiro, M. et al. Using Machine Learning to Improve the Prediction of Functional Outcome in Ischemic Stroke Patients. IEEE/ACM Trans. Comput. Biol. Bioinform. 15, 1953–1959 (2018)
  • Zeleňák K, Krajina A, Meyer L, et al. How to Improve the Management of Acute Ischemic Stroke by Modern Technologies, Artificial Intelligence, and New Treatment Methods. Life. 2021;11(6):488. doi:10.3390/life11060488
  • Alawieh A, Zaraket F, Alawieh MB, Chatterjee AR, Spiotta A. Using machine learning to optimize selection of elderly patients for endovascular thrombectomy. Journal of NeuroInterventional Surgery. 2019;11(8):847-851. doi:10.1136/neurintsurg-2018-014381
  • Shih C, Chu W, Chang Y-W. The Causes Analysis of Ischemic Stroke Transformation into Hemorrhagic Stroke using PLS (partial Least Square)-GA and Swarm Algorithm. undefined. Published 2019. Accessed February 4, 2022.
  • Garg R, Prabhakaran S, Holl J, Faigle R, Naidech A. Abstract TP366: Using Machine Learning to Predict Tracheostomy After Intracerebral Hemorrhage. Stroke. 2019;50(Suppl_1). doi:10.1161/str.50.suppl_1.tp366
  • Ge Y, Wang Q, Wang L, et al. Predicting post-stroke pneumonia using deep neural network approaches. International Journal of Medical Informatics. 2019;132:103986. doi:10.1016/j.ijmedinf.2019.103986
  • Chauhan S, Vig L, De Filippo De Grazia M, Corbetta M, Ahmad S, Zorzi M. A Comparison of Shallow and Deep Learning Methods for Predicting Cognitive Performance of Stroke Patients From MRI Lesion Images. Frontiers in Neuroinformatics. 2019;13. doi:10.3389/fninf.2019.00053
  • Bochniewicz EM, Emmer G, McLeod A, Barth J, Dromerick AW, Lum P. Measuring Functional Arm Movement after Stroke Using a Single Wrist-Worn Sensor and Machine Learning. Journal of Stroke and Cerebrovascular Diseases. 2017;26(12):2880-2887. doi:10.1016/j.jstrokecerebrovasdis.2017.07.004
  • Lansberg, M. G. et al. RAPID automated patient selection for reperfusion therapy: a pooled analysis of the Echoplanar Imaging Thrombolytic Evaluation Trial (EPITHET) and the Diffusion and Perfusion Imaging Evaluation for Understanding Stroke Evolution (DEFUSE) study. Stroke doi:10.1161/STROKEAHA.110.609008
  • Foroushani HM, Hamzehloo A, Kumar A, et al. Accelerating Prediction of Malignant Cerebral Edema After Ischemic Stroke with Automated Image Analysis and Explainable Neural Networks. Neurocritical Care. Published online August 20, 2021. doi:10.1007/s12028-021-01325-x
  • Semprini M, Laffranchi M, Sanguineti V, et al. Technological Approaches for Neurorehabilitation: From Robotic Devices to Brain Stimulation and Beyond. Frontiers in Neurology. 2018;9. doi:10.3389/fneur.2018.00212
  • Campbell BCV, Ma H, Ringleb PA, et al. Extending thrombolysis to 4·5–9 h and wake-up stroke using perfusion imaging: a systematic review and meta-analysis of individual patient data. The Lancet. 2019;394(10193):139-147. doi:10.1016/s0140-6736(19)31053-0 ‌
notion image
notion image
notion image