Background: Artificial intelligence (AI) is being adopted rapidly across healthcare. In stroke care, AI-enabled clinical decision support software is increasingly becoming part of the infrastructure. As these technologies mature, concerns remain about “black box” artificial intelligence models whose decision-making processes are not readily understood. This abstract aims to improve transparency by presenting the artificial intelligence techniques supporting a novel radiofrequency-based neurodiagnostic device.
Methods: The emu™ Brain Scanner (EMVision, Australia) is a bedside device employing AI-based models to analyse measured dielectric properties of the brain to detect hidden signatures to support stroke diagnosis. A model was developed using a proprietary database containing emu™ scans acquired from patients with suspected stroke. Self-supervised learning is used to identify generalisable structures within the signals; the learned feature representations are then used for classification by comparing an unseen case with cases of known diagnosis. This approach ground predictions to known examples that have undergone expert review.
Results: Algorithms developed in this manner demonstrated promising preliminary performance for both haemorrhage (92% sensitivity, 85% specificity) and ischemia (95% sensitivity, 80% specificity) detection algorithms. Further data collection in patients with suspected acute stroke is ongoing, next generation AI models and techniques, together with a validation study intended to support regulatory submission.
Conclusions: Carefully developed artificial intelligence techniques can achieve promising results in complex stroke datasets and are being applied in a new generation of portable brain scanner technology. Approaches that produce clinically interpretable results improve transparency, helping to build clinical trust and support adoption.