[4L1-GS-10] AI application: Medicine / Healthcare

[4L1-GS-10-02] Cerebral Artery Occlusion Inference Using Pulse Waves: Selection of Autoencoder for Occlusion Existence Classification

〇Kenshi Ohzono¹, Riko Fujisawa¹, Miho Ohsaki¹, Kimiaki Shirahama ¹, Mami Matsukawa¹, Yasuyo Kobayashi², Kozue Saito², Hiroshi Yamagami³ (1. Doshisha University, 2. Nara Medical University, 3. National Hospital Organization)

Keywords:Cerebral Artery Occlusion, Pulse Wave, Autoencoder

Cerebral artery occlusion poses the risk of death or severe sequelae, and prompt diagnosis and treatment after its onset determine the prognosis. Hence, we have been developing an occlusion diagnosis support system that can be used in emergency medical services. The system consists of a pulse wave measurement device and an occlusion inference method. The method infers if there is an occlusion using pulse waves possibly containing the reflection from the occlusion. The difficulty in defining features and the scarcity of occlusion cases make occlusion inference challenging. We propose a method that automatically extracts features and performs an occlusion classification using cases with no occlusion. In this method, an autoencoder (AE) reconstructs input pulse waves, and thresholding the reconstruction performance classifies the existence of occlusion. Since there are various choices for AE, the MLP, RNN (RNN, LSTM, GRU), and CNN AEs were compared and selected in experiments. As a result, regarding all the AEs, the proposed method achieved higher correct classification rates and F-values than those of random and all-positive classifications. GRU and CNN had comparatively higher performances. The effectiveness of our unsupervised classification method was demonstrated, where especially GRU AE and CNN AE were contributable.