https://orcid.org/0000-0002-8539-6408
Figures
Abstract
Background
Mechanical thrombectomy (MT) eligibility for acute ischemic stroke (AIS) patients depends upon clinical and advanced imaging assessments like CT perfusion (CTP). Assessment complexities and limited access to advanced imaging investigations are known challenges. We developed machine-learning models using routinely collected clinical and imaging data to predict MT eligibility.
Methods
Age, National-Institutes-of-Health-Stroke-Scale-Score (NIHSS), last-known-well-time (LKWT), noncontrast-CT (NCCT) scan and CT-angiography (CTA) report from consecutive cohort of 260 AIS-suspected patients treated at a stroke centre during Apr’20 to Dec’23 were retrospectively collected. 160 underwent MT for anterior-circulation large vessel occlusion (LVOa); rest were MT ineligible. MT eligibility was determined based on clinical and imaging investigations including CTP during routine-care. The dataset was split into train:test sets (50:50 split). A commercially available artificial-intelligence algorithm calculated infarct volume and ASPECT score (ASPECTSq) from the NCCTs. We developed two supervised models using Gradient-Boosting-Machines. MODEL1 utilized age, NIHSS, LKWT, ASPECTSq and infarct volume as inputs; MODEL2 additionally included the presence/absence of LVOa as input. The target/response variable used for our supervised learning methods was whether the patients were MT eligible or not as determined during routine-care. Performance of the models were investigated using the test set.
Results
Among 130 patients (mean age ± standard-deviation: 67.4 ± 14.2 years; 61 males) in test set, 80 (61.5%) were MT eligible; rest were ineligible. The area-under-the-receiver-operating-characteristics-curve, sensitivity and specificity of MODEL1 were 0.76 (95% CI: 0.67–0.85), 85% (75.6–91.2) and 60% (46.2–72.4), respectively. They were 0.92 (0.88–0.96), 82.5% (72.7–89.3) and 82% (69.2–90.2), respectively, for MODEL2.
Citation: Agarwal A, Bharti N, Ghosh T, Golla S, Bains NK, Chamadia R, et al. (2025) Development and internal validation of multimodal machine learning models for predicting eligibility for mechanical thrombectomy in suspected stroke patients using routinely collected clinical and imaging data. PLoS One 20(10): e0334242. https://doi.org/10.1371/journal.pone.0334242
Editor: Aldobrando Broccolini, Fondazione Policlinico Universitario Agostino Gemelli IRCCS, ITALY
Received: May 19, 2025; Accepted: September 24, 2025; Published: October 10, 2025
Copyright: © 2025 Agarwal et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data and the Python code used to train and test the multi-modal XGBoost prediction models reported in this paper are available in https://github.com/arjunagarwal899/multimodal-machine-learning-models-for-predicting-mechanical-thrombectomy-eligibility. Imaging data (non-contrast head CT scans) is not required for model training and testing. If access to imaging data is required, please send a request to corresponding author and access will be granted subject to institutional requirements.
Funding: The study received funding from Qure.ai. The funder was involved in the study design, data collection and analysis, decision to publish and in the preparation of the manuscript.
Competing interests: AA, NB, TG, SG, RC, DR and PP are employees of Qure.ai. NKB and AIQ declared that no competing interests exist. This does not alter our adherence to PLOS ONE policies on sharing data and materials.
Introduction
Mechanical thrombectomy (MT) is established as a treatment of choice in acute ischemic stroke (AIS) patients with anterior circulation large vessel occlusion (LVOa) [1]. Whilst initially MT was recommended for patients with LVOa within 6 hours of last known well time (LKWT) and with Alberta Stroke Program Early Computed Tomography Score (ASPECTS) ≥ 6, randomized controlled trials have demonstrated benefit of MT in eligible patients within 6–24 hours of LKWT, with consequent recommendations in the guidelines [1–3]. Eligibility for MT in patients presenting during the extended window (within 6–24 hours of LKWT) is extrapolated from randomized controlled trials [2,3] utilizing imaging modalities such as Computed Tomography Perfusion (CTP) or Diffusion-weighted Magnetic Resonance Imaging (DW-MRI), while for those presenting within 6 hours of LKWT, eligibility for MT can be determined based on non-contrast CT (NCCT) brain and CT angiogram (CTA) or MRI/MR angiogram alone [1]. Recently, another randomized controlled trial has shown superior effectiveness of MT in large core infarct (ASPECTS score ≤ 5) LVOa patients presenting within 6.5 hours of LKWT compared to medical care alone [4].
While CTP and DW-MRI imaging are largely used in advanced stroke care centers for MT eligibility in suspected AIS patients, these advanced imaging modalities have limited availability in under-resourced settings, and require qualified personnel to interpret imaging studies and post-processing software to generate CTP parameter maps which are further complicated by variability in parameter maps generated by different commercially available software packages [5–7]. Though the utilization of CTA and CTP for suspected stroke patients are on the rise, data suggests that they are still underutilized [8]. As a consequence, many otherwise eligible patients may not undergo MT or experience delay in undergoing MT which can adversely affect their outcomes. It is thus pertinent to investigate if eligibility for MT in AIS suspect patients can be accurately predicted by using data without using advanced imaging investigations. Such models could potentially assist concerned healthcare professionals in clinical decision making, especially resource constrained settings. While numerous pre-hospital models exist in literature, most are limited by being unimodal, focusing on predicting LVOa likelihood rather than MT eligibility [9]. Additionally, these models often have suboptimal inter-rater reliability and require complex assessments that are challenging for pre-hospital personnel to implement [9].
In this study, we report development and internal validation of machine learning models using multimodal data from routinely collected patient data and imaging modalities such as NCCT and/or CTA that would predict MT eligibility.
Materials and methods
Data collection
For the development and validation of the multimodal prediction model (s), we retrospectively collected anonymized data from 260 consecutive stroke-suspect patients aged 18 years and older admitted to an advanced stroke care center in the United States of America (Zeenat Qureshi Stroke Institute, St. Cloud, Minnesota, USA) between April 2020 and December 2023. All the 260 patients underwent stroke evaluation with National Institutes of Health Stroke Scale (NIHSS) [10], NCCT, CTA and CTP and had a LKWT of maximum 24 hours. All imaging investigations were performed within 24 hours of symptoms onset. The radiologist interpretations of CTA and CTP parameter maps from a CTP post-processing software package were in the collected data, but not CTA and CTP scans. NCCT scans were available. 160 (61.5%) of the patients underwent MT with eligibility determined by clinical, imaging and CTP perfusion parameters. Of the remaining 100 (38.5%) patients who did not undergo MT, 30 were found ineligible for MT despite presence of LVOa and the remaining 70 did not have a LVOa. Patients with missing data on age, NIHSS and imaging (NCCT, CTA and CTP) data were excluded.
Model development
Training data.
We used data from 130 patients (80 who underwent MT [MT eligible]) and 50 who did not) for training the multimodal prediction model and reserved the remaining 130 (50:50 train:test split) as testing dataset (Fig 1). The target/response variable used for our supervised learning was based on whether the patients underwent MT (MT eligible, n = 80) or not (MT ineligible, n = 50).
Predictor variables.
We developed two supervised binary classification prediction models, named MODEL1 and MODEL2, using the eXtreme Gradient Boosting (XGBoost) machine learning algorithm [11]. Predictor variables of MODEL1 included five predictor variables – 1) age, 2) LKWT, 3) NIHSS score, 4) infarct volume in millilitre (ml) predicted by an existing deep learning based segmentation model (qER, Qure.ai) from NCCT scan [12–15] and 5) ASPECT score (ASPECTSq) predicted by qER from NCCT scan (Fig 2). MODEL2 included an additional binary predictor variable for the presence or absence of LVOa based on the CTA interpretation of the original reporting neuroradiologist. Details of the predictor variables based on NCCT scans (infarct volume and ASPECTSq) are detailed below.
ASPECTS: Alberta Stroke Program Early Computed Tomography Score.
Infarct volume:
We utilized an existing acute infarct segmentation model (qER; Manufacturer: Qure.ai) to determine infarct volume. This model is a semantic segmentation-based deep learning framework that employs a transformer-based encoder architecture, paired with a robust segmentation head. The encoder is pretrained on 125,000 head NCCT scans to classify the presence of infarcts, distinguishing between acute and chronic infarcts. The segmentation head leverages the encoder’s feature representations to generate three target masks: all infarcts, acute infarcts, and chronic infarcts. The final output is a refined segmentation map that accurately delineates infarcted regions. For our analysis, we focused on the volume of the acute infarct mask, as our focus was on acute stroke cases.
ASPECTSq:
qER’s ASPECT score model combines two components: the infarct segmentation model and the ASPECT anatomy model. The ASPECT anatomy model predicts a mask of the 10 regions associated with ASPECT scoring, while the infarct segmentation model (described in section above) predicts the acute infarct mask (Fig 3). When integrated, these models provide an efficient means to quickly obtain the ASPECT score (a value between 0 and 10) for any NCCT scan. For reporting purpose, we call this ASPECT score outputted from qER based on processing the NCCT scan as ASPECTSq. The ASPECT anatomy model was trained on an extensive dataset of 50,000 head NCCT scans.
XGBoost models
Our multimodal prediction models (MODEL1 and MODEL2) were built using XGBoost algorithm. XGBoost is a widely used powerful ensemble machine learning algorithm known for its high performance in classification tasks, particularly in scenarios involving complex and heterogeneous data, such as in medical decision-making [11,16]. By utilizing an ensemble of decision trees, XGBoost can capture non-linear relationships and interactions between the predictor variables, making it an ideal choice for our model’s classification task. XGBoost is reported to produce accuracies which are comparable to many state of the art machine learning algorithms, including deep neural networks [16].
In our study, the XGBoost algorithm was trained to model the probability for MT eligibility based on the predictor variables in the training data described earlier. Multiple hyperparameters of the XGBoost model, such as learning rate, tree depth, minimum child weight, subsample ratio and feature sampling were optimized through grid search to find the best hyperparameter values. The final model configuration was selected based on optimal MT eligibility area under the receiver operating characteristics curve (AUROC) performance. A fixed random state was used to ensure reproducibility across experiments.
Model testing and statistical analysis
A sample size of 130 patients (80 MT eligible and 50 MT ineligible) would be enough to estimate an AUROC of 0.75 or more with minimum 9% precision (half-width of 95% CI) [17]. The fine-tuned MODEL1 and MODEL2 were evaluated using the testing set consisting of data from 130 patients (80 who underwent MT [MT eligible]) and 50 who did not). The reference standard (ground truth) used for evaluation was based on whether the patients were deemed MT eligible or not during their assessment at the stroke care center. AUROC, sensitivity and specificity of the models are reported along with their 95% confidence intervals. To calculate sensitivity and specificity, an optimal threshold maximizing sensitivity and specificity was applied to the probability score for each model in order to classify the instances into MT eligible or not. We also report accuracy (proportion of correctly classified instances) and F1 score (harmonic mean of positive predictive value and sensitivity). We also performed the feature (predictor variable) importance analysis and SHapley Additive exPlanations (SHAP) [18] analysis of predictor variables.
Since MODEL1 and MODEL2 also used infarct volume and ASPECTSq as predictor variables, which were in turn based on prediction outputs from an existing deep learning model (qER), we performed a comparison of qER’s acute infarct prediction with radiologists as an exploratory analysis. For this analysis, all 260 head NCCT scans (training and testing data combined) were independently reviewed by four radiologists (one neuroradiologist and three general radiologists, 7–15 years range of experience) blinded to any other information to discern the presence/absence of acute infarct. The sensitivity and specificity of the radiologists in detecting acute infarct were descriptively compared with that of qER.
Ethical considerations
The study proposal was reviewed by Institutional Review Board (IRB) at University of Missouri-Columbia (IRB Project Number: 2095636) and was deemed exempt for review. Informed consent was not required due to the retrospective nature of the study. Only the authorized investigators (NKB and AIQ) from the institution had access to identifiable information during data collection. Any identifiable information was permanently removed immediately after data collection and only de-identified data was used for model development and validation. Data were accessed during February 2023 to March 2024 in accordance with annual IRB exemption requirements.
Results
Model training
Baseline characteristics of the 130 patients included in the training data are shown in Table 1. Best performing models after hyperparameter tuning had 10 decision trees for MODEL1 and 12 for MODEL2 and a maximum depth of 3 for both the models, preventing overfitting while maintaining sufficient complexity. A minimum child weight of 2 to control tree splitting, a subsample ratio of 0.6 to introduce randomness for better generalization, and 60% feature sampling to reduce feature correlation were also fine-tuned through grid search for both models. Learning rate was found to be optimal at 0.15 for MODEL1 and for 0.10 for MODEL2. The best checkpoint yielded an AUROC of 0.86 (95% CI: 0.80–0.92) for MODEL1 and 0.96 (95% CI: 0.93–0.99) for MODEL2.
Multimodal prediction model performance on testing data
The mean age of the patients in the testing data was 67.4 years (standard deviation: 14.2 years) and 61 (46.9%) were males. The baseline characteristics of patients in the testing data are detailed in Table 2.
MODEL1, which incorporates clinical data and results from NCCT scans, achieved an AUROC of 0.76 (95% CI: 0.67–0.85) (Table 3). In contrast, MODEL2, which integrates clinical data with results from both NCCT and CTA (presence of LVOa), demonstrated an AUROC of 0.92 (95%CI: 0.88–0.96) (Fig 4). The performance of both the models was better in patients with LKWT of less than 6 hours (AUROC: 0.78 [95% CI: 0.67–0.88] for MODEL1 and 0.95 [95% CI: 0.90–0.99] for MODEL2) compared to that in patients with LKWT of 6–24 hours (AUROC: 0.65 [95% CI: 0.47–0.82] for MODEL1 and 0.86 [95% CI: 0.73–0.95] for MODEL2). The improvement in AUROC of MODEL2 compared to MODEL1 is reflected further with a larger improvement in specificity (82% for MODEL2 and 60% for MODEL1) than improvement in sensitivity (82.5% for MODEL2 and 85% for MODEL1). The feature importance analysis showed that NIHSS score and infarct volume were the top two predictor variables in MODEL1 whereas in MODEL2 these were the presence of LVOa and NIHSS (Fig 4). The top predictor variable as per SHAP analysis was NIHSS score for MODEL1 (S1 Fig) and presence of LVOa for MODEL2 (S2 Fig) which were consistent with the feature importance. However, the ranking of the other predictor variables was not similar in order between feature importance analysis and SHAP analysis. This could be because feature importance analysis considers how much each feature contributes to model predictions overall and SHAP analysis considers how much each feature contributes to prediction of each individual instance.
Comparison of acute infarct detection in head NCCT scans between qER and radiologists
Of the 260 patients, 190 (73.1%; 160 [84.2%] MT eligible and 30 [15.8%] MT ineligible) had confirmed presence of LVOa and thus were considered to have presence of acute infarct. The rest of the 70 patients had no LVOa and no perfusion deficit and thus were considered as controls (no acute infarct) in this exploratory analysis. Overall, the sensitivity of the radiologists ranged from 5.0% to 21.2% whereas that of qER was found to be 25.0% (S1 Table and S3 Fig). The specificity of the radiologists ranged from 78.6% – 98.6% and that of qER was found to be 85.7%.
Discussion
In this study, we trained multimodal prediction models using clinical and imaging data from head NCCT and CTA scans to predict MT eligibility in a patient with AIS. Using only clinical data and basic imaging investigation (NCCT), the model achieved a good AUROC of 0.76 (95% CI: 0.67–0.85) in testing set. When neuroradiologist’s interpretation on the presence or absence of LVOa from head CTA was added as a new predictor variable, the model had significant improvement in performance as indicated by an excellent AUROC of 0.92 (95% CI: 0.88–0.96). At optimal threshold, the specificity of MODEL2 was much higher than that of MODEL1, but the sensitivities of the two models were not that largely different.
The benefit of MT is known to be time dependent. For every 15-minute reduction in door-to-reperfusion time during MT, an estimated 39 out of 1,000 patients show improved outcomes at 3 months [19]. The economic and clinical implications of accurate MT eligibility prediction are substantial. Unnecessary delays in treatment can significantly increase neurological disability, healthcare costs, and long-term patient burden. Delay in transferring patients from primary centres to comprehensive stroke care centres and requirement of CTP may contribute to delay in access to MT [20,21]. Our models could potentially be useful for suspected AIS patients, especially those presenting initially in low-resource settings without advanced imaging infrastructure such as CTP or DW-MRI, enabling faster referral to higher stroke care centers. The predictor variables in our models would already be routinely collected as part of basic assessment of stroke suspect patients. In a real-world setting, these models could act as triaging tools for enabling faster referrals. Since false negative (incorrectly classifying an MT eligible patient as ineligible) results are likely to cause worse outcomes for patients than false positive results, a high sensitivity is preferable for such models even at the expense of specificity. These models can be technically integrated into the hospital information technology infrastructure so that real-time model prediction results can be viewed by concerned healthcare professionals such as emergency care physicians, radiologists, neurologists and/or neurointerventional professionals, either within hospital or remotely.
To our knowledge, this is the first multimodal MT eligibility prediction model of its kind described in the literature. There are several unimodal prediction models available to assess the likelihood for LVOa in pre-hospital setting using basic clinical data alone [9]. However, not all LVOa patients are eligible for MT. Koster et al. conducted a systematic review to assess the effectiveness of several such models in determining the eligibility for MT and observed that the AUROCs of these models ranged from 0.75 to 0.83 [22]. Interestingly, NIHSS score is reported to have an AUROC of 0.81 which is comparatively higher than the AUROC of our MODEL1 which also included NIHSS score as one of the predictor variables. NIHSS score was the top predictor variable in MODEL1. A single center data and the relatively small size of the training dataset in our study may be the reason for the lower AUROC in our MODEL1. However, the addition of a new predictor variable indicating presence/absence of LVOa significantly improved model performance to 0.92 which is comparatively higher than the best performing pre-hospital clinical prediction model (FAST-ED). FAST-ED, however, suffers from suboptimal interrater reliability and uses complex items which are difficult for pre-hospital care personnel to assess [23,24]. Seetge et al. used a logistic regression model using age, NIHSS score at admission, and pre-morbid modified Rankin Scale score as predictor variables to stratify AIS patients into low-, moderate-, and high-risk groups, to guide treatment decisions on thrombolysis, MT, combination therapy (thrombolysis + MT), or standard care, and observed that the model had an AUROC of 0.86 in predicting 90-day outcomes [25]. Alwood et al. used CTP parameter map outputs from two CTP post-processing software packages and reported that there were significant differences in core and penumbra volume estimates between the outputs of software packages [26]. They also found that there were no statistically significant differences between the two software packages when they were used to determine MT eligibility based on the DEFUSE-3 [3] eligibility criteria. Our models incorporate simple clinical data which are routinely collected in hospitals and outputs from imaging data which can be fed into the model in real-time when scans are acquired and/or reported. We acknowledge that MODEL2, however, requires input from a head CTA study which may not be available in many low-resource settings and that comparison of pre-hospital clinical prediction models with in-hospital prediction models may not be very ideal.
In our exploratory analysis of comparing acute infarct detection performance of radiologists with our existing deep learning model (qER) in NCCT scans, we observed better acute infarct detection rates for qER. The sensitivity of both radiologists and qER were comparatively higher in NCCT scans from patients who were ineligible for MT despite LVOa. These patients were ineligible for MT due to established hypodensity on NCCT suggestive of established infarct and thus might have been easier to detect. The sensitivity in detecting AIS from NCCT scans is generally considered to be on the lower side with one study reporting it to be about 26% [27] and in general considered to be around 40% [28]. It is noteworthy that the sensitivity of qER in detecting acute infarct was 40% in MT eligible patients and 60% in MT ineligible (but with presence of LVOa) patients.
Major limitations of our study include using a single center data, relatively smaller sample size in the training data and use of testing data from the same source as that of the training data. In future work, we intend to retrain models using data from multiple sources and use independent external data sources for performing external validation to improve the generalizability of the models. We only experimented with XGBoost machine learning algorithm, and it can be argued that other machine learning algorithms such as, but not limited to, random forests and logistic regression, could also have been experimented with. We chose XGBoost in this study because it is well known to be amongst the best performing machine learning algorithms. For MODEL2, we ideally could have utilized another existing head CTA LVOa prediction algorithm using CTA scans as input to generate the predictor variable for the presence of LVOa, but we did not have access to CTA imaging data and thus had to rely on reporting neuroradiologist’s interpretation as the predictor variable.
Conclusions
Our multimodal prediction models show potential in predicting the eligibility for MT using routinely collected clinical and imaging data during assessment of a suspected stroke patient. Such models can be useful in stroke management, especially in low-resource settings and can enable faster referrals of potentially eligible patients to higher stroke care centers. Further research with multicenter data and external validation is required to provide further corroborative evidence in this regard.
Supporting information
S1 Table. Acute infarct detection performance by radiologists and qER in noncontrast head CT scans.
https://doi.org/10.1371/journal.pone.0334242.s001
(DOCX)
S1 Fig. MODEL1 predictors and their mean SHAP values.
SHAP: SHapley Additive exPlanations.
https://doi.org/10.1371/journal.pone.0334242.s002
(TIFF)
S2 Fig. MODEL2 predictors and their mean SHAP values.
SHAP: SHapley Additive exPlanations.
https://doi.org/10.1371/journal.pone.0334242.s003
(TIFF)
S3 Fig. Noncontrast CT head acquired from a suspected acute ischemic stroke patient for whom a mechanical thrombectomy was done at the site.
CT perfusion (not shown in the figure) indicated 70 ml of tissue with cerebral blood flow (CBF) less than 30% and 133 ml of tissue with time to maximum (Tmax) of more than 6 seconds suggesting substantial amount (mismatch ratio: 1.9) of potentially salvageable tissue over the left middle cerebral artery territory. In the retrospective investigation, the acute infarct was detected by qER (drawn in red contours on the figure), but not by any of the four radiologists in the noncontrast CT scan.
https://doi.org/10.1371/journal.pone.0334242.s004
(TIFF)
Acknowledgments
We acknowledge the contributions of Polamoni Ranjana Devi for project conceptualization and management, and Avneesh Jha, Abhipray Dohane and Amit Chouksey for facilitating the annotation process for noncontrast head CT scans. We also want to acknowledge Pashmeen Lakhani from Zeenat Qureshi Stroke Institute for her contributions during study conduct.
References
- 1. Powers WJ, Rabinstein AA, Ackerson T, Adeoye OM, Bambakidis NC, Becker K, et al. Guidelines for the Early Management of Patients With Acute Ischemic Stroke: 2019 Update to the 2018 Guidelines for the Early Management of Acute Ischemic Stroke: A Guideline for Healthcare Professionals From the American Heart Association/American Stroke Association. Stroke. 2019;50(12):e344–418. pmid:31662037
- 2. Nogueira RG, Jadhav AP, Haussen DC, Bonafe A, Budzik RF, Bhuva P, et al. Thrombectomy 6 to 24 Hours after Stroke with a Mismatch between Deficit and Infarct. N Engl J Med. 2018;378(1):11–21. pmid:29129157
- 3. Albers GW, Marks MP, Kemp S, Christensen S, Tsai JP, Ortega-Gutierrez S, et al. Thrombectomy for Stroke at 6 to 16 Hours with Selection by Perfusion Imaging. N Engl J Med. 2018;378(8):708–18. pmid:29364767
- 4. Costalat V, Jovin TG, Albucher JF, Cognard C, Henon H, Nouri N, et al. Trial of Thrombectomy for Stroke with a Large Infarct of Unrestricted Size. N Engl J Med. 2024;390(18):1677–89.
- 5. Konstas AA, Goldmakher GV, Lee T-Y, Lev MH. Theoretic basis and technical implementations of CT perfusion in acute ischemic stroke, part 1: Theoretic basis. AJNR Am J Neuroradiol. 2009;30(4):662–8. pmid:19270105
- 6. Katyal A, Bhaskar SMM. Value of pre-intervention CT perfusion imaging in acute ischemic stroke prognosis. Diagn Interv Radiol. 2021;27(6):774–85. pmid:34792033
- 7. Koopman MS, Berkhemer OA, Geuskens RREG, Emmer BJ, van Walderveen MAA, Jenniskens SFM, et al. Comparison of three commonly used CT perfusion software packages in patients with acute ischemic stroke. J Neurointerv Surg. 2019;11(12):1249–56. pmid:31203208
- 8. Kim Y, Lee S, Abdelkhaleq R, Lopez-Rivera V, Navi B, Kamel H, et al. Utilization and Availability of Advanced Imaging in Patients With Acute Ischemic Stroke. Circ Cardiovasc Qual Outcomes. 2021;14(4):e006989. pmid:33757311
- 9. Nguyen TTM, van den Wijngaard IR, Bosch J, van Belle E, van Zwet EW, Dofferhoff-Vermeulen T, et al. Comparison of Prehospital Scales for Predicting Large Anterior Vessel Occlusion in the Ambulance Setting. JAMA Neurol. 2021;78(2):157–64. pmid:33252631
- 10.
NIH. Stroke Scale. [cited 7 Feb 2025]. Available from: https://www.ninds.nih.gov/sites/default/files/documents/NIH-Stroke-Scale_updatedFeb2024_508.pdf
- 11. Chen T, Guestrin C. XGBoost: {A} Scalable Tree Boosting System. CoRR. 2016;abs/1603.0. Available from: http://arxiv.org/abs/1603.02754
- 12. Chiramal JA, Johnson J, Webster J, Nag DR, Robert D, Ghosh T, et al. Artificial Intelligence-based automated CT brain interpretation to accelerate treatment for acute stroke in rural India: An interrupted time series study. PLOS Glob Public Health. 2024;4(7):e0003351.
- 13. Hillal A, Sultani G, Ramgren B, Norrving B, Wassélius J, Ullberg T. Accuracy of automated intracerebral hemorrhage volume measurement on non-contrast computed tomography: a Swedish Stroke Register cohort study. Neuroradiology. 2023;65(3):479–88. pmid:36323862
- 14. Chilamkurthy S, Ghosh R, Tanamala S, Biviji M, Campeau NG, Venugopal VK, et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. Lancet. 2018;392(10162):2388–96. pmid:30318264
- 15. Pettet G, West J, Robert D, Khetani A, Kumar S, Golla S, et al. A retrospective audit of an artificial intelligence software for the detection of intracranial haemorrhage used by a teleradiology company in the United Kingdom. BJR Open. 2024;6(1):tzae033. pmid:39479271
- 16. Moore A, Bell M. XGBoost, A Novel Explainable AI Technique, in the Prediction of Myocardial Infarction: A UK Biobank Cohort Study. Clin Med Insights Cardiol. 2022;16:11795468221133611. pmid:36386405
- 17. Zhou XH, Obuchowski NA, McClish DK. Statistical Methods in Diagnostic Medicine. 2011, p. 1–545.
- 18. Rodríguez-Pérez R, Bajorath J. Interpretation of machine learning models using shapley values: application to compound potency and multi-target activity predictions. J Comput Aided Mol Des. 2020;34(10):1013–26. pmid:32361862
- 19. Saver JL, Goyal M, van der Lugt A, Menon BK, Majoie CBLM, Dippel DW, et al. Time to Treatment With Endovascular Thrombectomy and Outcomes From Ischemic Stroke: A Meta-analysis. JAMA. 2016;316(12):1279–88. pmid:27673305
- 20. Aroor SR, Asif KS, Potter-Vig J, Sharma A, Menon BK, Inoa V, et al. Mechanical Thrombectomy Access for All? Challenges in Increasing Endovascular Treatment for Acute Ischemic Stroke in the United States. J Stroke. 2022;24(1):41–8. pmid:35135058
- 21. Prabhakaran S, Ward E, John S, Lopes DK, Chen M, Temes RE, et al. Transfer delay is a major factor limiting the use of intra-arterial treatment in acute ischemic stroke. Stroke. 2011;42(6):1626–30. pmid:21527756
- 22. Koster GT, Nguyen TTM, van Zwet EW, Garcia BL, Rowling HR, Bosch J, et al. Clinical prediction of thrombectomy eligibility: A systematic review and 4-item decision tree. Int J Stroke. 2019;14(5):530–9. pmid:30209989
- 23. Lima FO, Silva GS, Furie KL, Frankel MR, Lev MH, Camargo ÉCS, et al. Field Assessment Stroke Triage for Emergency Destination: A Simple and Accurate Prehospital Scale to Detect Large Vessel Occlusion Strokes. Stroke. 2016;47(8):1997–2002. pmid:27364531
- 24. Kesinger MR, Sequeira DJ, Buffalini S, Guyette FX. Comparing National Institutes of Health Stroke Scale among a stroke team and helicopter emergency medical service providers. Stroke. 2015;46(2):575–8. pmid:25538202
- 25. Seetge J, Cséke B, Karádi ZN, Bosnyák E, Szapáry L. Stroke-SCORE: Personalizing Acute Ischemic Stroke Treatment to Improve Patient Outcomes. J Pers Med. 2025;15(1):18. pmid:39852210
- 26. Alwood BT, Meyer DM, Ionita C, Snyder KV, Santos R, Perrotta L, et al. Multicenter comparison using two AI stroke CT perfusion software packages for determining thrombectomy eligibility. J Stroke Cerebrovasc Dis. 2024;33(7):107750. pmid:38703875
- 27. Chalela JA, Kidwell CS, Nentwich LM, Luby M, Butman JA, Demchuk AM, et al. Magnetic resonance imaging and computed tomography in emergency assessment of patients with suspected acute stroke: a prospective comparison. Lancet. 2007;369(9558):293–8. pmid:17258669
- 28. Brazzelli M, Sandercock PA, Chappell FM, Celani MG, Righetti E, Arestis N, et al. Magnetic resonance imaging versus computed tomography for detection of acute vascular lesions in patients presenting with stroke symptoms. Cochrane Database Syst Rev. 2009;(4):CD007424. pmid:19821415