SeLaNet: A Noninvasive Distilled Neural Network to Predict the Dense SEC/LAAT Based on Multi-layer Perceptron

• Published in: International Conference on Advanced Cloud and Big Data pp. 112 - 117
• Main Authors: Zhang, Ziheng, Gu, Fei, Zhang, Zeyang, Wang, Xinyu, Sun, Shikun, Zhou, Jingya
• Format: Conference Proceeding
• Language: English
• Published: IEEE 28.11.2024
• Subjects: Accuracy; Clinical diagnosis; dense spontaneous echo contrast and left atrial appendage thrombus; Echocardiography; Logistic regression; multi-layer perceptron; multi-teacher knowledge distillation; Nearest neighbor methods; Neural networks; Random forests; Support vector machines; Surgery; Training
• ISSN: 2573-301X
• DOI: 10.1109/CBD65573.2024.00030

Dense spontaneous echo contrast (SEC) or left atrial appendage thrombus (LAAT) is common in patients with atrial fibrillation (AF) or atrial flutter (AFL). Although LAAT detachment is an important cause of ischemic stroke, Dense SECILAAT is difficult to accurately predict in order to meet the surgical prediction criterion. Transesophageal echocardiography is one of the few approaches to acquire correct results, but is an invasive test. In this paper, we propose SeLaN et, a distilled neural network to predict and diagnose dense SECILAAT, based on 55-dimensional clinical characteristics associated with dense SECILAAT. SeLaNet is based on a multi layer perceptron (MLP). Multi-teacher knowledge distillation is performed with three screened distant teachers and hidden layer loss. We also compared SeLaN et with some machine learning methods including k-nearest neighbor (KNN), logistic regression, random forest and support vector machine (SVM). SeLaNet achieved an accuracy of 95.96%, and an AUC-ROC of 0.9765. Such performance can play a unique role in clinical diagnosis.