The recording of Monday’s Sano Seminar is now available

During the seminar, Rosmary Blanco from the Computational Neuroscience Team delivered a talk entitled “Toward Trustworthy AI for Acute Ischemic Stroke Triage with EEG: From Signal Processing to Uncertaint -Aware Decision Support.”

The presentation focused on the development of trustworthy AI systems for early triage of acute ischemic stroke using EEG, with a particular emphasis on pre-hospital settings such as ambulances and on-scene emergency care.

Key takeaways from the talk

Project goal: to develop a trustworthy AI system for early ischemic stroke triage based on EEG, suitable for use in pre-hospital conditions.

Clinical data were collected within the ElectroStroke and AI Stroke clinical trials at Amsterdam UMC, using 8-channel EEG with dry electrodes, recorded by paramedics.

A major challenge was EEG signal quality, affected by motion artifacts, environmental noise, and the lack of a universal processing pipeline for low-density EEG.

A systematic comparison of EEG artifact removal methods was performed, including ASR, Wavelet Transform, EMD, and hybrid approaches, using simulated, benchmark, and real clinical data.

Wavelet Transform performed best for general denoising of low-channel EEG, while EMD better preserved pathological patterns such as sharp waves and rhythmic discharges.

Automatic machine learning (AutoML – AutoPrognosis) was used to classify ischemic stroke vs. stroke mimics, enabling fair and reproducible comparisons between processing pipelines.

Although standard performance metrics (accuracy, AUC) were similar across methods, clinical error analysis revealed meaningful differences from a patient’s safety perspective.

In pre-hospital triage, sensitivity is critical; a hybrid EMD + Wavelet approach minimized the number of missed strokes.

The system incorporated uncertainty quantification (both aleatoric and epistemic) and a rejection mechanism to refer uncertain cases to a physician.

A cascaded model architecture was proposed to mirror clinical reasoning:
Is this a stroke? → Is it an LVO? → Is the stroke still treatable?, while accounting for the temporal dynamics of EEG.

Watch the full seminar recording on YouTube: