Prediction of Functional Outcome After Acute Ischemic Stroke: Comparison of the CT-DRAGON Score and a Reduced Features Set

In what universe do you live in where you think survivors give a rats' ass about predicting less than 100% recovery? I'm really curious as to your thinking about how this is going to help survivors.

Prediction of Functional Outcome After Acute Ischemic Stroke: Comparison of the CT-DRAGON Score and a Reduced Features Set

Anouk Lesenne^1,2, Jef Grieten^2,3, Ludovic Ernon⁴, Alain Wibail⁴, Luc Stockx⁵, Patrick F. Wouters¹, Leentje Dreesen⁵, Elly Vandermeulen², Sam Van Boxstael², Pascal Vanelderen^2,6, Sven Van Poucke², Joris Vundelinckx², Sofie Van Cauter⁵ and Dieter Mesotten^2,6*

• ¹Department of Anesthesiology and Perioperative Medicine, Ghent University Hospital, Ghent, Belgium
• ²Department of Critical Care Services, Ziekenhuis Oost-Limburg Genk, Genk, Belgium
• ³Department of Anesthesiology, VU University Amsterdam, Amsterdam, Netherlands
• ⁴Department of Neurology, Ziekenhuis Oost-Limburg Genk, Genk, Belgium
• ⁵Department of Medical Imaging, Ziekenhuis Oost-Limburg Genk, Genk, Belgium
• ⁶UHasselt, Faculty of Medicine and Life Sciences, Diepenbeek, Belgium

Background and Purpose: The CT-DRAGON score was developed to predict long-term functional outcome after acute stroke in the anterior circulation treated by thrombolysis. Its implementation in clinical practice may be hampered by its plethora of variables. The current study was designed to develop and evaluate an alternative score, as a reduced set of features, derived from the original CT-DRAGON score.

Methods: This single-center retrospective study included 564 patients treated for stroke, in the anterior and the posterior circulation. At 90 days, favorable [modified Rankin Scale score (mRS) of 0–2] and miserable outcome (mRS of 5–6) were predicted by the CT-DRAGON in 427 patients. Bootstrap forests selected the most relevant parameters of the CT-DRAGON, in order to develop a reduced set of features. Discrimination, calibration and misclassification of both models were tested.

Results: The area under the receiver operating characteristic curve (AUROC) for the CT-DRAGON was 0.78 (95% CI 0.74–0.81) for favorable and 0.78 (95% CI 0.72-0.83) for miserable outcome. Misclassification was 29% for favorable and 13.5% for miserable outcome, with a 100% specificity for the latter. National Institutes of Health Stroke Scale (NIHSS), pre-stroke mRS and age were identified as the strongest contributors to favorable and miserable outcome and named the reduced features set. While CT-DRAGON was only available in 323 patients (57%), the reduced features set could be calculated in 515 patients (91%) (p<0.001). Misclassification was 25.8% for favorable and 14.4% for miserable outcome, with a 97% specificity for miserable outcome. The reduced features set had better discriminative power than CT-DRAGON for both outcomes (both p<;0.005), with an AUROC of 0.82 (95% CI 0.79–0.86) and 0.83 (95% CI 0.77–0.87) for favorable and miserable outcome, respectively.

Conclusions: The CT-DRAGON score revealed acceptable discrimination in our cohort of both anterior and posterior circulation strokes, receiving all treatment modalities. The reduced features set could be measured in a larger cohort and with better discrimination. However, the reduced features set needs further validation in a prospective, multicentre study.

Clinical Trial Registration: http://www.clinicaltrials.gov. Identifiers: NCT03355690, NCT04092543.

Introduction

With an incidence of 14 million patients annually, ischemic stroke is the second largest cause of death globally after ischemic heart disease. It has a major burden of morbidity as well, with an estimated annual 52 million disability-adjusted life years (1). Prognostic tools that predict outcome of acute ischemic stroke potentially provide early identification of patients who are likely to have a good or poor outcome despite treatment. If this tool has a high specificity for miserable outcome, it could be helpful in counseling patients and relatives, because estimations of outcome in stroke patients remain largely subjective at this moment. Moreover, and probably more applicable, these scores could be used for case-mix adjustments for benchmarking purposes.

In this view, several prognostic scoring systems have been developed to address this need, such as the ASTRAL, the CT-DRAGON, the iSCORE and the PLAN score (2). However, they have not been widely implemented in clinical practice, due to several limitations. First, a large number of input variables is required, some of which are hard to determine in the acute setting. Second, these scores are often tailored to subpopulations of stroke patients, depending on the localization of the stroke and/or the treatment received. Third, there is still a lack of validation in large patient populations using real world data, which are more prone to missingness and inaccuracies.

Treatment options for acute stroke have strongly evolved over the last decade. Most scoring systems were developed in the era of thrombolysis, while thrombectomy and its combination with thrombolysis only have been implemented over the last years (3–9). These developments certainly should have influenced reliable estimates of outcome and long-term effect of treatment.

To enhance implementation of prognostic tools, Fahey et al. advised to validate existing prognostic tools in different patient populations and treatment settings, opposed to designing new ones (10). Notably the CT-DRAGON score (Dense Artery, modified Rankin Scale, Age, Glucose, Onset-to-Treatment and NIHSS) has already been validated in previous studies and adapted to different diagnostics and treatments in stroke patients (11–17). The MRI-DRAGON score was developed and externally validated to deal with patients, in whom MRI was used as the first-line diagnostic tool (18, 19). Recently, the DRAGON score has also been modified to deal with patients, undergoing mechanical thrombectomy (20). These modifications are relevant since decision to perform mechanical thrombectomy in patients with wake-up strokes is often based on MRI diffusion-weighted imaging (DWI). Despite the modifications to specific subpopulations of stroke patients, its potential remains underutilized because of the missing data. A “light version” of the DRAGON score with less input variables, selected by machine learning, might thus be an alternative.

Machine learning has many applications, among which prediction of outcomes in healthcare. Machine learning techniques consist of algorithms able to solve problems by learning from experiences. Mathematical models are built and trained by providing training data. When new data are supplied, the models are able to generalize their learned expertise and make accurate predictions. Dimensionality reduction is a more recent application of machine learning. This process strives to reduce the number of variables under consideration. By “feature selection” input variables from existing scoring systems can be reduced and optimized (21, 22). With a reduced features set dynamic predictive models can be built, deployed and monitored over time.

We aimed to validate the CT-DRAGON score in all ischemic stroke localizations and for all treatment options, including a conservative treatment policy. The predictive power was then compared with a model, that included a set of the individual components of the CT-DRAGON score, selected by machine learning techniques.