Abstract WP221: Feasibility of an AI-assisted transcranial duplex sonography protocol for early detection of intracerebral hemorrhage: the HYPER-AI-SCAN single-center prospective study

Abstract

Background: Early identification of intracerebral hemorrhage (ICH) is crucial in the hyperacute stroke setting, yet computed tomography (CT) may be delayed or unavailable in prehospital contexts. We evaluated the feasibility of standardized transcranial duplex sonography (TCD) combined with a model-agnostic AI pipeline for early ICH screening. Methods: This prospective single-center study was conducted at Hospital Universitario Vall d’Hebron (Barcelona, Spain) in adult patients with acute stroke. Following CT confirmation, participants underwent portable transtemporal TCD using a standardized 5s sweep (depth 19–20 cm, tilt ≈25°). Videos were de-identified, preprocessed, and sampled with a first-pass policy (20 frames from the initial 2.5 s). A linear classifier was fine-tuned on encoded features from a pre-trained, publicly available ultrasound encoder model and a lightweight attention-based multiple-instance learning head was trained on 5-fold cross-validation with patient-disjoint splits (within each fold, patients were partitioned 44/10/14 into train/validation/test) to aggregate the 20 frames per sweep into a single patient-level probability of suspected ICH. Feasibility endpoints included acquisition time, acoustic window quality, and AI diagnostic signal. Results: Sixty-eight patients were enrolled (baseline in Table 1). Median onset-to-TCD time was 26.5 h [IQR 16.8–40.4]. The examination required 21.5 s [IQR 12–31.3] from probe positioning to diagnostic images. Acoustic window quality was good in 54.4%, average in 30.9%, and poor/absent in 14.7%. Frame AI readout with the first-pass policy achieved mean balanced accuracy 0.55 ± 0.06 (sensitivity 0.21 ± 0.16, specificity 0.89 ± 0.11), F1 0.33± 0.12. The patient-level readout yielded: balanced accuracy 0.56 ± 0.09 (sensitivity 0.60 ± 0.38, specificity 0.52 ± 0.40), and F1 0.48 ± 0.17. Conclusions: HYPER-AI-SCAN demonstrates proof-of-concept feasibility for rapid TCD-based ICH screening within clinically realistic timeframes. The very short acquisition time underscores its suitability for integration into acute stroke pathways.While patient-level performance remains modest at this phase, the technical workflow is viable and provides a foundation for refinement on larger cohorts toward clinical implementation.

Ähnliche Arbeiten