Cross-Entropy Learning for Aortic Pathology Classification of Artificial Multi-Sensor Impedance Cardiography Signals

Tobias Spindelböck, Sascha Ranftl*, Wolfgang von der Linden

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

Abstract

An aortic dissection, a particular aortic pathology, occurs when blood pushes through a tear between the layers of the aorta and forms a so-called false lumen. Aortic dissection has a low incidence compared to other diseases, but a relatively high mortality that increases with disease progression. An early identification and treatment increases patients’ chances of survival. State-of-the-art medical imaging techniques have several disadvantages; therefore, we propose the detection of aortic dissections through their signatures in impedance cardiography signals. These signatures arise due to pathological blood flow characteristics and a blood conductivity that strongly depends on the flow field, i.e., the proposed method is, in principle, applicable to any aortic pathology that changes the blood flow characteristics. For the signal classification, we trained a convolutional neural network (CNN) with artificial impedance cardiography data based on a simulation model for a healthy virtual patient and a virtual patient with an aortic dissection. The network architecture was tailored to a multi-sensor, multi-channel time-series classification with a categorical cross-entropy loss function as the training objective. The trained network typically yielded a specificity of (93.9±0.1)% and a sensitivity of (97.5±0.1)%. A study of the accuracy as a function of the size of an aortic dissection yielded better results for a small false lumen with larger noise, which emphasizes the question of the feasibility of detecting aortic dissections in an early state.
Original languageEnglish
Article number1661
JournalEntropy
Volume23
Issue number12
DOIs
Publication statusPublished - 2021

Keywords

  • Aortic dissection
  • Aortic pathology
  • Convolutional neural network
  • Cross-entropy
  • Data fusion
  • Impedance cardiography
  • Machine learning
  • Time-series classification

ASJC Scopus subject areas

  • Information Systems
  • Electrical and Electronic Engineering
  • Mathematical Physics
  • Physics and Astronomy (miscellaneous)

Fingerprint

Dive into the research topics of 'Cross-Entropy Learning for Aortic Pathology Classification of Artificial Multi-Sensor Impedance Cardiography Signals'. Together they form a unique fingerprint.

Cite this