Can we trust AI in Acute Stroke Care?
A patient arrives at the emergency department with symptoms of an acute ischemic stroke (AIS). Within minutes, AI software analyzes the brain scan and flags a possible large vessel occlusion, helping the stroke team prioritize treatment. But as AI becomes increasingly integrated into stroke care, an important question remains: how much should clinicians rely on its recommendations?
Artificial intelligence has made remarkable progress in acute stroke care over the past decade. AI-powered clinical decision support systems (CDSS) can now assist clinicians by automatically detecting large vessel occlusions, estimating infarct core and penumbra volumes from perfusion imaging, and providing standardized Alberta Stroke Program Early CT Score (ASPECTS) assessments to evaluate early ischemic changes. By rapidly processing complex imaging data, these tools have the potential to shorten the time to diagnosis and treatment, which is an important advantage in a condition where every minute counts. However, there are certain hurdles that remain before these tools can be safely integrated into routine clinical practice.
A recent systematic review identified 121 studies proposing AI-based clinical decision support systems (CDSS) for predicting clinical outcomes in acute ischemic stroke (AIS) and found that many of these models had critical threats to their validity. Nearly half of the studies (42%) did not use an independent test set to evaluate their models, but instead tested the models on the validation dataset used during model development. This suggests that the reported performance may be overestimated and may not generalize well to real-world clinical settings. Another major concern is the lack of demographic transparency. Only 3% of the analyzed studies reported the race of the patients used to train the AI models, and none reported socioeconomic status. Without this information, there is a greater risk of bias, as the models may not generalize well to diverse populations. Furthermore, for AI to be useful in emergency settings, it must be data-efficient. The sources point out that many current models require over 10 clinical variables or multiple imaging sequences, which can impede the fast-paced clinical workflow. To address these challenges, the authors recommend involving clinicians early in the development process and following standardized reporting frameworks, such as the Minimum Information for Medical AI Reporting (MINIMAR) checklist, to improve transparency and facilitate the evaluation and comparison of AI systems.
Finding the right balance of trust in an AI system is critical for clinicians, as both undertrust and overtrust can negatively affect patient care. Even if an AI system correctly identifies a large vessel occlusion in most patients, variations in imaging quality, uncommon vascular anatomy, or atypical stroke presentations can lead to incorrect predictions. In these scenarios, AI should support rather than replace human expertise. One proposed solution to address these challenges is the use of explainable artificial intelligence (XAI). Unlike traditional black-box algorithms that provide only a final prediction, explainable AI seeks to show how that prediction was reached. For example, an AI tool could highlight the specific brain regions that influenced its detection of a vessel occlusion or indicate which imaging features contributed most to estimating infarct volume. By making AI recommendations more transparent, clinicians may be better equipped to evaluate whether they align with the patient’s clinical presentation and imaging findings.
Ultimately, the goal is not to maximize trust in AI, but to foster appropriate trust. Clinicians need to understand both the strengths and the limitations of AI systems so they can confidently incorporate AI into decision-making while recognizing situations that require additional scrutiny. Therefore, it is crucial to design AI technologies that fit naturally into clinical workflows. Projects such as MDR in AIS are helping to ensure that AI becomes a trusted clinical partner, enhancing rather than replacing human expertise.
Author: Liva Araka, Rīga Stradiņš University
References:
- Akay EMZ, Hilbert A, Carlisle BG, et al. Artificial Intelligence for Clinical Decision Support in Acute Ischemic Stroke: A Systematic Review. Stroke. 2023;54(6):1505–1516. https://doi.org/10.1161/STROKEAHA.122.041442
- Zihni E, McGarry BL, Kelleher JD. Moving Toward Explainable Decisions of Artificial Intelligence Models for the Prediction of Functional Outcomes of Ischemic Stroke Patients. In: Digital Health. Exon Publications; 2022. https://doi.org/10.36255/exon-publications-digital-health-explainable-decisions
Keywords:
acute ischemic stroke (AIS) # AI-powered clinical decision support systems (CDSS) # large vessel occlusion # demographic transparency # explainable artificial intelligence (XAI)