matthiaszeller / vad-0d-dnn

Code for paper "Deep Neural Network to Accurately Predict Left Ventricular Systolic Function Under Mechanical Assistance", Bonnemain et al.

VAD-0D-DNN

About

This repository contains the code for the paper Deep Neural Network to Accurately Predict Left Ventricular Systolic Function Under Mechanical Assistance, Jean Bonnemain, Matthias Zeller, Luca Pegolotti, Simone Deparis, Lucas Liaudet.

Requirements

• Matlab R2020a
• Openmodelica 1.16.0
• Python 3
  • Standard packages: matplotlib, pandas, numpy
  • OMPython
  • Keras

Project Overview

The project is subdivided in the following sections:

• 0D Model Implementation in modelica/original/Mathcard.mo: 0D Model of the cardiovascular system implemented in the Modelica language.

• 0D Model Simulation in Simulation_script: Python code running the simulations with OMPython package

  • Generate a dataset by solving the model with different parameter values
  • Perform a sensitivity analysis: fix 3 heart-failure parameters and vary the 4th one
  • Case study: simulate a few scenarios by varying heart failure severity, pump speed

• Preprocessing in Pre-processing: Python code

  • Extract systemic arterial pressure and pulmonary arterial pressure from simulation data
  • Compute Fourier coefficients
  • Generate DNN input and output files

• DNN Training and Testing in Deep_learning: Python code

  • Build and format dataset for the DNN

• Post-processing: Post-processing, Matlab code

  • Compute hemodynamic quantities on the test set

• Project pipelining: pipelining Python code gluing data generation, preprocessing, DNN training

  • script_pipeline.py: Automatic way to generate datasets with different pump configurations, merge them, train different DNNs on the merged dataset
  • script_check_sim_data.py: if an error occured while generating simulation data with script_pipeline.py, this script helps you to identify the corrupted parts and prevents from recomputing everyting
  • script_test_dnn.py : once the best DNN is chosen, use this script to run simulations on test data with this chosen DNN

Reproductibility

The procedure described below was tested on Ubuntu 20.04. The settings as set in script_pipeline.py generated about 2.1 TB of data and the whole procedure took about 20h to complete.

Overview

1. Set the range of LVAD rotational speeds, the simulation settings (e.g. sampling), the number of Fourier coefficients to retain, the DNN architectures (and many other parameters) in pipeline/script_pipeline.py

2. Set sampling settings in Pre-processing/setup_preprocessing.py

3. Run the script pipeline/script_pipeline.py, which

   1. Generates simulation data for each specified pump configuration
   2. Merges data of all configurations in a single .mat file and creates training and test sets
   3. Trains the specified DNNs

   Note 1: this script might take several hours (or even days) to run

   Note 2: the output folder will have the following contents:

   .
   ├── 4000_LVAD_AP
   │   ├── outputs
   │   │   ├── Ursino1998Model_VAD2_output_0.mat
   │   │   ├── Ursino1998Model_VAD2_output_1.mat
   │   │   ├── ...
   │   ├── parameters.txt
   │   ├── X.mat
   │   ├── Y.mat
   │   └── ...
   ├── ...
   ├── 6000_LVAD_AP
   │   ├── ...
   ├── dataset.mat
   ├── dnns
   │   ├── dnn_10_layers_128_neurons
   │   │   ├── DNN_Performance.eps
   │   │   ├── history.bin
   │   │   ├── losses.eps
   │   │   ├── model.h5
   │   │   ├── Ytestpred.txt
   │   │   └── Ytest.txt
   │   ├── dnn_10_layers_16_neurons
   │   │   ├── ...
   │   ├── ...
   │   └── normdata.mat
   ├── logs.log
   └── mainlog.log
   

4. Once you identified a DNN architecture that performs well, run pipeline/script_test_dnn.py to generate simulations on test data

5. Compute hemodynamic quantities on test data:

   1. Setup paths in Post-processing/setupproj.m to target the outputs/ folder of the selected DNN architecture
   2. Run test_dnnmodelevaluation.m