REDEFINING the position of AI in acute stroke care, a presentation delivered in the course of the session ‘AI in Hospital Neurology’ on the European Academy of Neurology (EAN) Congress 2026 examined how AI is evolving past automated picture evaluation to develop into a medical decision-support software. Susanna Wegener, Senior Doctor, Division of Neurology, College Hospital Zurich, Switzerland, offered rising analysis spanning prognostic prediction, therapy decision-making, and secondary stroke prevention. The presentation delivered a key message that profitable AI implementation will rely not solely on more and more refined algorithms, but additionally on clinically significant knowledge, clear predictions, and the belief of the clinicians utilizing them.
PREDICTING FUNCTIONAL RECOVERY AFTER STROKE
A central theme within the medical makes use of of AI centres round whether or not it will probably meaningfully enhance prognostic evaluation following acute stroke. To discover this, Wegener offered a examine evaluating skilled stroke neurologists with convolutional neural community (CNN) fashions in predicting modified Rankin Scale (mRS) scores 3 months post-stroke.1 Specialists have been supplied with medical knowledge, MRI imaging, or a mixture of each. Utilizing medical info alone, neurologists achieved an accuracy of roughly 60%, with CNN fashions performing comparably. Including MRI knowledge improved prognostic accuracy for each clinicians and AI; nevertheless, the CNN mannequin persistently outperformed professional neurologists total, highlighting its potential to reinforce, reasonably than exchange, medical decision-making.
The findings have been validated utilizing CT angiography knowledge from the MR CLEAN trial,2 the imaging modality extra generally utilized in routine medical follow. The examine in contrast prognostic efficiency between neurologists, neurologists assisted by the MR PREDICT rating, and AI basis fashions educated on CT knowledge. As within the earlier examine, neurologists achieved an accuracy of roughly 60–65%, whereas the AI fashions persistently demonstrated superior efficiency. Though the MR Predict rating improved prognostic accuracy amongst much less skilled neurologists, it provided little further profit for extremely skilled clinicians (Herzog et al., unpublished knowledge, 2026). Wegener famous that MR PREDICT scores depend on handbook evaluation of imaging variables, together with the Alberta Stroke Program Early CT Rating, collateral standing, and occlusion location. Appreciable variability exists amongst stroke specialists when assigning these measures, whereas AI fashions can routinely extract imaging options immediately from CT scans, decreasing subjectivity whereas bettering the consistency of prognostic evaluation.
A key discovering from this examine confirmed that neurologists predicted extra beneficial purposeful outcomes than sufferers in the end achieved, whereas neither the AI mannequin nor the MR Predict rating did. Clinicians overestimated the advantages of reperfusion therapies, whereas components reminiscent of older age and male intercourse have been weighted too negatively. These findings spotlight the affect of cognitive biases on prognostic evaluation and reinforce the potential for AI to supply extra goal resolution help when built-in alongside medical experience.
BUILDING BETTER AI MODELS
Wegener argued that noticed limitations in AI might mirror deficiencies within the final result measures used to coach present fashions. The mRS rating supplies a broad evaluation of incapacity however doesn’t distinguish stroke-related incapacity from unrelated occasions. Consequently, sufferers who die from unrelated stroke causes are nonetheless assigned the worst rating. Growing extra significant, goal endpoints, reminiscent of quantitative motor assessments, might allow AI fashions to higher seize therapy results and generate extra clinically related predictions. As Wegener summarised, “it relies on what you feed in. What you get out is what you feed in,” underscoring that AI efficiency is basically restricted by the standard and relevance of the information used to coach it.
Recognising that mannequin efficiency relies on the standard of enter knowledge, Wegener highlighted variables often ignored by present prediction fashions. Within the potential STOP-Stroke examine, clinicians recorded their predicted prognosis and the components influencing their selections (STOP-Stroke, unpublished knowledge, Westphal). Preliminary knowledge from 50 sufferers confirmed that earlier than imaging, prognosis was pushed primarily by Nationwide Institutes of Well being Stroke Scale (NIHSS) rating, age, and onset-to-door time. Imaging contributed lower than anticipated as clinicians’ assessments continued to depend on medical knowledge. Earlier than discharge, frailty, cognitive decline, and former stroke emerged as further key predictors. Collectively, these findings present consideration of things that affect medical selections are integral to growing refined AI fashions.
EXPANDING AI BEYOND ACUTE CARE
Past acute stroke administration, Wegener highlighted AI’s potential to strengthen secondary stroke prevention by figuring out sufferers with beforehand undetected atrial fibrillation. As many sufferers have imaging options suggestive of cardioembolic stroke regardless of no documented arrhythmia, extended cardiac monitoring is commonly required earlier than anticoagulation might be prescribed.
To handle this problem, Wegener mentioned analysis utilizing UK Biobank knowledge, the place machine studying fashions mixed cardiac imaging with structural markers of atrial cardiomyopathy to establish people vulnerable to growing atrial fibrillation (Deseoe et al., in press, 2026). It was discovered that the ratio of left atrial to left ventricular quantity outperformed atrial dimension alone, with these findings subsequently validated in stroke cohorts. Comparable predictive efficiency was achieved utilizing basis fashions educated on routine ECGs obtained throughout regular sinus rhythm, outperforming blood biomarkers and traditional medical danger scores throughout a number of validation cohorts, together with UK Biobank, Brazilian main care, stroke cohorts, and Holter ECG datasets.
Though potential medical trials are nonetheless required to determine whether or not these approaches enhance affected person outcomes, the findings recommend that AI can detect delicate indicators of future atrial fibrillation utilizing checks which are already routinely carried out, enabling earlier identification of high-risk sufferers and extra well timed secondary stroke prevention.
FUTURE DIRECTIONS
Trying forward, the profitable integration of AI into routine stroke care will rely on greater than advances in algorithm efficiency. Growing clinically significant final result measures, incorporating ignored prognostic variables, and validating fashions throughout numerous populations will likely be important. Supporting this effort, Wegener highlighted the European MAGIC Consortium, a multicentre initiative that brings collectively anonymised medical and stroke imaging datasets from throughout Europe.3 By creating giant, numerous datasets for AI growth and validation, the consortium goals to enhance mannequin robustness and speed up functions starting from prognostic prediction to therapy choice and secondary stroke prevention. Potential medical trials will then be required to find out whether or not these applied sciences enhance medical decision-making and affected person outcomes.
