Letter by Goyal and Ospel Regarding Article, “Multimodal Predictive Modeling of Endovascular Treatment Outcome for Acute Ischemic Stroke Using Machine-Learning”

HomeStrokeVol. 52, No. 2Letter by Goyal and Ospel Regarding Article, "Multimodal Predictive Modeling of Endovascular Treatment Outcome for Acute Ischemic Stroke Using Machine-Learning" Free AccessLetterPDF/EPUBAboutView PDFView EPUBSections ToolsAdd to favoritesDownload citationsTrack citationsPermissions ShareShare onFacebookTwitterLinked InMendeleyRedditDiggEmail Jump toFree AccessLetterPDF/EPUBLetter by Goyal and Ospel Regarding Article, "Multimodal Predictive Modeling of Endovascular Treatment Outcome for Acute Ischemic Stroke Using Machine-Learning" Mayank Goyal, MD, PhD and Johanna M. Ospel, MD Mayank GoyalMayank Goyal https://orcid.org/0000-0001-9060-2109 Department of Clinical Neurosciences (M.G., J.M.O.), University of Calgary, Canada. Department of Radiology (M.G.), University of Calgary, Canada. and Johanna M. OspelJohanna M. Ospel Department of Clinical Neurosciences (M.G., J.M.O.), University of Calgary, Canada. Department of Radiology, University Hospital of Basel, Switzerland (J.M.O.). Originally published25 Jan 2021https://doi.org/10.1161/STROKEAHA.120.033159Stroke. 2021;52:e83–e84To the Editor:We congratulate Brugnara et al1 to their thoughtfully designed and meticulously executed study, in which they assessed the performance of multimodal, machine-learning based outcome prediction in an acute ischemic stroke patient sample that underwent endovascular treatment. There is no doubt that machine learning techniques will play a key role in stroke outcome prediction models of the future, and this vanguard study constitutes an excellent starting point in this regard.With their article, Brugnara et al teach us several lessons that are crucial for physicians involved in acute stroke care, namely:Accuracy of outcome prediction at the time of decision making, that is, based on pretreatment clinical variables and conventional imaging alone is limited. The area under the curve (AUC) for the baseline model in the study was 0.740, which ranges in between the predictive accuracy of a coin flip (AUC of 0.5) and a perfect model (AUC of 1).2 Of note, the AUC of this baseline model is probably a rather optimistic estimate, since variables such as hyperlipidemia, HbA1c, and glucose levels, which were part of the model, are rarely available at the time of decision making.Advanced imaging does not improve outcome prediction. When CT perfusion-derived ischemic core and penumbra volumes and the penumbra-to-core-ratio were added to the baseline model, no significant improvement was seen (AUC of 0.747). As the authors mention in their discussion, this should cause us to critically re-evaluate the supposed benefit of advanced imaging for outcome prediction and endovascular treatment decision making.As expected, retrospective outcome prediction, that is, outcome prediction with a model that incorporates pretreatment, treatment and post-treatment variables, yields the highest predictive performance (AUC of 0.856), but it is still not perfect. It would be of great interest to investigate which factors explain the remaining variability in outcomes that cannot be explained by our current outcome prediction models.Importantly, among the top 5 predictors in the highest performing retrospective model, 2 were post-treatment variables, namely 24-hour National Institutes of Health Stroke Scale score and 24-hour infarct volume. Since they are unknown at the time presentation, these variables are of no utility for clinicians who have to decide whether to treat a given patient or not. The only treatment variable among the top 5 predictors was groin-puncture-to-recanalization time, which constitutes an important starting point for in-hospital workflow improvement endeavors.By far the most important predictor, however, was the premorbid functional status of a patient. This seems intuitively logical: a patient with substantial co-morbidities may have less recovery capacities and is more prone to post-treatment complications, which ultimately results in worse functional outcomes. However, the expectations of such patients with regard to treatment and poststroke outcome might also be different, something that cannot easily be captured in prediction models. The only pretreatment imaging variable among the top 5 predictors was ischemic core volume on baseline noncontrast with a rather low relative importance of 8%, which should serve us as a reminder for what imaging in acute stroke actually is: a tiny part of the big picture.AcknowledgmentsDrs Ospel and Goyal contributed in conceptualization, drafting, and critical revision of the article.Sources of FundingNone.Disclosures Dr Goyal reports personal fees from Medtronic, Stryker, Microvention and Mentice outside the submitted work; and a patent to Systems of acute stroke diagnosis. The other author reports no conflicts.FootnotesFor Disclosures, see page e84.References1. Brugnara G, Neuberger U, Mahmutoglu MA, Foltyn M, Herweh C, Nagel S, Schönenberger S, Heiland S, Ulfert C, Ringleb PA, et al.. Multimodal predictive modeling of endovascular treatment outcome for acute ischemic stroke using machine-learning.Stroke. 2020; 51:3541–3551. doi: 10.1161/STROKEAHA.120.030287LinkGoogle Scholar2. Hosmer D, Lemeshow S, Sturdivant R. Applied Logistic Regression. Wiley; 2013.CrossrefGoogle Scholar Previous Back to top Next FiguresReferencesRelatedDetails February 2021Vol 52, Issue 2Article InformationMetrics © 2021 American Heart Association, Inc.https://doi.org/10.1161/STROKEAHA.120.033159PMID: 33493061 Originally publishedJanuary 25, 2021 PDF download Advertisement SubjectsIschemic StrokeQuality and OutcomesRevascularizationTreatmentVascular Disease