Fra Robin

Gå til: navigasjon, søk



In 2005 Elle et al. developed a novel technique at The Intervention Centre at the Oslo University Hospital for detecting cardiac ischemia using a 3-axis accelerometer attached to the human heart. The intended goal is to develop a continuous monitoring device that can detect complications during and after a cardiac surgery. In this thesis I will try to develop an algorithm to automate the classification task using deep learning architectures.

Thesis Structure

  • Introduction
  • Background
    • The Human Heart
      • The anatomy of the heart
      • The physiology of the heart
      • Cardiac dysfunctions
      • Monitoring techniques
    • Deep Learning
      • Traditional neural networks
      • Convolutional neural networks
      • Recurrent neural networks
    • Time-Series Signals
    • Classifiers And Regressors
    • Signal Processing Theory
    • Available Data
  • Experiment 1: Domain Analysis
    • Idea
    • Data preparation
      • Time feature independent data set
      • Time feature dependent data set
    • Implementation
    • Results
  • Experiment 2: Classifying Cardiac Heart Functions
    • Idea
    • Preparations
    • Implementation
    • Results
  • Experiment 3: Image Classification
    • Idea
    • Preparations
    • Implementation
    • Results
  • Comparison To Other Research
  • Future Work
  • Conclusion


Is it possible to use motion data to predict the condition of the human heart? If so, can Deep Learning be used, and is it effective?

Do we need an invasive method for classification (accelerometer), or is it sufficient with a non-invasive method (ECG)?


Experiment 1


Given an input signal represented in a specific domain, an output signal represented in another domain, and a neural network that transforms the input signal to the output signal, the more features the input-signal contains than the output-signal, the better the transformation should be if both signals represent the same real-world features.


Plot demonstrating the R-peak feature occurring at arbitrary time steps given a sample division of 250 samples for each training sample. Features can also appear several times for each training sample, as seen in the second training sample in the ECG plot:

One of the first regressors going form accelerometer data to ECG data. This dense regressor is predicting an ECG signal when provided an ACC signal. The data is re-sampled to 250 Hz to reduce training time:

When constructing conv. network, I could let me be inspired by [this] paper.

Experiment 2

Experiment 3

Possible representation of data:

Personlige verktøy