In current stroke guidelines, advanced imaging techniques hold a crucial role in the decision-making process for treatment triage of Large Vessel Occlusion (LVO). Among these advanced imaging modalities, Computed Tomography Angiography (CTA) is a pivotal tool. It serves not only for evaluating treatment eligibility but also for assessment of arterial collateral supply and predicting functional stroke outcome prognosis. It has also been proven that CTA is more sensitive in detecting early infarction signs compared to non-contrast Computed Tomography (CT) (Camargo et al., 2007). Infarct cores determined on CTA images are strongly correlated with lesions defined on diffusion-weighted MRI (DWI) scans (Schramm et al., 2004). Recent studies have also revealed CTA’s potential to provide valuable information for long-term prognostication (van Seeters et al., 2015; Sallustio et al., 2017).

Recently, the emergence of artificial intelligence (AI) models has opened new possibilities for the prediction of long-term outcome from baseline stroke imaging information. These models enable the extraction of prognostic information directly from admission CTA scans, offering the potential to forecast patients’ outcomes. Unlike previous approaches that relied on manually engineered imaging biomarkers or complex preprocessing steps (Avery et al., 2022; Zhang et al., 2023), our approach streamlines the prediction process by taking the 3D images as input with minimal preprocessing in the end-to-end automated pipeline that uses raw unprocessed CTA scan to generate long-term outcome predictions. The disadvantage of complex preprocessing is the risk of information loss and the risk of introducing unknown biases. Deep learning makes it possible to keep the preprocessing steps small and at the same time preserve biomarkers that are currently unknown and not easily visually discernible.

In this study, we trained and tested separate deep learning models to predict 3-month outcomes after LVO thrombectomy from admission CTA scans with and without additional treatment and clinical variables. We compared models’ performance in independent test cohort and analyzed model biases. Such tools can facilitate objective prognostication of LVO stroke patients in acute setting. The prognostication of stroke outcomes by the presented models is especially useful for situations with the absence of reliable neurological exams and provides support for informed discussion regarding outcomes with patients and family members as well as establishing long-term goals of stroke management.

Our study demonstrates that by utilizing admission CTA scans, an automated model can accurately predict 3-month outcomes after thrombectomy in patients with LVO stroke. The end-to-end pipeline, which involves the preprocessing of admission CTAs, can be smoothly integrated into the clinical workflow. Furthermore, although not achieving statistical significance, we showed that the addition of post-thrombectomy reperfusion can increase the AUC of model predictions in the independent test cohort. The inclusion of clinical predictors such as age, sex, and neurological severity alongside CTA data did not significantly enhance model performance. Hence, CTA-based models can be especially useful in situations where clinical information is unreliable, such as in tele-stroke settings, language barriers, or when pre-existing morbidities make neurological exams challenging. It is worth noting that CTA scans are the de facto first line of imaging used to screen and diagnose LVO stroke, meaning that images are likely available before a complete clinical exam. Thus, a model that incorporates CTA data and accounts for reperfusion success can provide valuable prognostic information prior to the initial neurological exam.

Recent clinical trials (RESCUE–Japan LIMIT, SELECT-2, ANGEL-ASPECT, and TENSION) have demonstrated improved functional outcomes after thrombectomy, even among patients with a large infarct core and anterior circulation LVO (Yoshimura et al., 2022; Bendszus et al., 2023; Huo et al., 2023; Sarraj et al., 2023). In this context, the focus of stroke imaging workflow should shift toward identifying patients with LVO and thus eligible for thrombectomy (regardless of infarct core estimates). Subsequently, it should also identify patients at higher risk of poor outcomes despite thrombectomy – or clinically ineffective reperfusion. These patients are potential candidates for post-thrombectomy treatments, such as neuroprotective or neuroregenerative therapies. Deep learning models, such as the one developed and validated in this study, can provide this critical information.

However, other groups have also applied deep learning models for the prognostication of stroke. One research group applied a “Structured Receptive Field Neural Network” (RFNN), a variant of the ResNet model, with emphasis on the advantages of RFNN models in the context of smaller training datasets (Hilbert et al., 2019). They reported an average AUC of 0.71 for their best-performing model. Despite the complexity of their model and the presumable advantage in training with small datasets, their results did not exceed ours. Notably, their model generated a single 2D image from CTA scans for the prediction task. Their group also provided no comparison with models containing admission clinical information (Hilbert et al., 2019).

Another study employs a specialized model architecture called a “Siamese network” to focus on hemispheric asymmetries (Oliveira et al., 2023). This approach utilizes a more complex model structure, but it has the advantage that images of varying qualities could be assessed more uniformly. This model type utilizes Non-Contrast CT (NCCT) images as the main input. The CTA images were only used to predict the presence or absence of stroke to use as one of the clinical inputs. For this, they were compressed to a 2D image using MIP (Maximum Intensity Projection). In the highest-performing version of this model, the research group achieved an AUC of 0.74 in a test set size of 60 patients. Notably, the group included patients both with and without LVO, whereas patients with LVO are far more likely to have worse outcomes (Smith et al., 2006), and thus the model could have differentiated between those with and without LVO; in contrast, we only included patients with LVO, therefore a more clinically homogenous patient cohort.

Furthermore, in another study the researchers extracted radiomics features from middle cerebral artery (MCA) regions of CTA scans. These were utilized for the prediction of 3-month outcomes in patients with LVO stroke (Avery et al., 2022). Comparing models with different inputs, “Radiomics + Treatment,” and “Radiomics + Treatment + Clinical,” achieved AUCs of 0.68, 0.74, and 0.82 in independent test cohort, respectively (Avery et al., 2022).

Yet, one of the strengths of our research is that we created a fully end-to-end automated preprocessing and prediction pipeline for anterior LVO strokes. Compared to the previously mentioned works, our preprocessing pipeline includes fewer steps and yet achieves the same or better results. For example, we avoided skull removal techniques for CTA images due to artifact risk, we utilized a pre-trained network to improve classification, and we refrained from using MIP preprocessed images as input to minimize data loss. The advantage of this is that we can reduce the error potential and bias of our algorithm. The ensemble model structure, in which several individually trained versions come to a joint decision, has not yet been reported for this task. Not only it minimizes the risk of overfitting, but it also allows the flexible addition of other clinical variables, without any negative impact on the performance of the image analysis. In our analysis, we compared how the model with clinical variables performed in contrast to the model without clinical variables. Another novelty of our method is the inclusion of thrombectomy success as a prognostic input. This variable can provide the best versus worst-case scenario prediction for complete reperfusion versus thrombectomy failure. The “CTA + Treatment” model holds great potential for prognostication in acute stroke settings, as it can provide a wide range of outcome predictions, from the least to the most favorable treatment results based on the presumed lowest to highest mTICI reperfusion. This approach enables the model to estimate the probabilities of 3-month outcomes based on potential thrombectomy success.

It’s worth noting that the model can solely rely on CTA scans, too, which is almost always available at the time of LVO diagnosis. Therefore, a model based on imaging information can provide rapid and objective predictions regardless of local expertise and inter-examiner variabilities.

In our work, we also compared ensemble models using different variables with each other and were able to do a comprehensive assessment of different input strategies. For example, we wanted to simulate situations in which the determination of NIHSS is not possible for various reasons (e.g., tele-stroke setting, pre-existing motor deficit due to musculoskeletal degenerative disease, or language barrier). Although the combination of age, NIHSS, sex, and treatment results emerge as a strong prognostic model (Cummock et al., 2023; Oliveira et al., 2023), a fully automated model based on imaging input alone can be useful for immediate risk stratification as soon as CTA scan is completed.

The use of an ensemble model as described by us also has the advantage of increasing generalizability. Since multiple versions of a model make a decision together by averaging each one’s probability, it results in less fluctuation between each patient and is less prone to overfit. Since it is widely known that deep learning models trained on a specific set of imaging characteristics may struggle to generalize to external datasets with different imaging properties (Li et al., 2023), we also took additional measures to avoid overfitting in the training process. For example, we applied data augmentation techniques as well as L2-regularization, as they are known to increase the generalizability of a trained model (Sanford et al., 2020; Li et al., 2023). Notably, prognostic models based on automated analysis of admission CTA can offer generalizable treatment guidance information given the widespread availability of CTA scans, even in rural areas. This additional information is provided regardless of the presence expert reviewers, and without need for additional advanced imaging (such as perfusion), extra radiation or contrast administration. Risk-stratification of patients can identify potential candidates for additional post-thrombectomy neuroprotective or neuroregenerative therapies.

Attention maps help visually illustrate the deep learning model’s perception and provide insight into the decision-making process of machine learning models. As depicted in Figure 4, cerebral areas within the MCA supply territory had the highest impact on the decisions of our models. These findings confirm that model predictions were based on attention to at-risk cerebral tissue on CTA scans of LVO stroke patients (Waqas et al., 2020).

Our study has limitations. In model bias analysis, we found that false predictions of our model more commonly involved patients with favorable clinical characteristics (Table 2). The class imbalance within the dataset, particularly within the lower mRS categories, may have exerted an influence on the model’s performance during both the training and testing phases. In theory, there are some strategies that can be applied in the future to mitigate the issue of overestimating classes. For example, generative adversarial networks (GANs) are increasingly applied to generate synthetic images that oversample minority classes (Islam and Zhang, 2020), therefore increasing the ability of the model to reliably detect minority classes. Other approaches are aimed at optimizing the analysis of the generated probabilities like prediction uncertainty (Zou et al., 2023) or using a conditional probability for bias correction in prediction (Alexandari et al., 2020).

Also, retrospective datasets may suffer from biases inherent in the data collection process. These biases could be related to patient selection, imaging protocols, or institutional practices. Moreover, the mTICI score included in the dataset were not from core laboratory and therefore subject to inter-examiner variability. In addition, stroke management and outcomes can evolve due to advancements in medical treatments, changes in clinical guidelines, or improvements in healthcare practices. Retrospective studies without external validation might not account for these temporal changes, affecting the model’s applicability to more recent patient cohorts.

We showed the feasibility of an end-to-end fully automated to predict post-thrombectomy outcomes from readily available admission CTA images and treatment data. The model can be practically useful in the absence of a reliable neurological exam which can also provide robust prognostication. The comparison of our work to existing literature suggests that a simple deep learning approach, as implemented in our model, strikes a pragmatic balance between performance, architectural simplicity, and preprocessing ease.

The data analyzed in this study is subject to the following licenses/restrictions: unknown. The datasets analyzed during the current study are available from the corresponding author on reasonable request. The programming code used for analysis can be found at: https://github.com/Fledermaus12/LVOstroke-DL. Requests to access these datasets should be directed to SP, sam.payabvash@yale.edu.

JS: Software, Writing – original draft, Writing – review & editing. FD: Formal analysis, Software, Writing – review & editing. TZ: Software, Writing – review & editing. AT: Software, Writing – review & editing. EA: Data curation, Writing – review & editing. AdM: Data curation, Writing – review & editing. AjM: Data curation, Writing – review & editing. CM: Data curation, Writing – review & editing. GF: Data curation, Writing – review & editing. VT-L: Data curation, Writing – review & editing. SA: Data curation, Software, Writing – review & editing. JD: Data curation, Writing – review & editing. LS: Data curation, Writing – review & editing. KS: Data curation, Writing – review & editing. SP: Conceptualization, Funding acquisition, Methodology, Project administration, Resources, Supervision, Writing – original draft, Writing – review & editing.