Information-theoretic methods for food supply network identification in foodborne disease outbreaks

Abigail L. Horn1,2, Marcel Fuhrmann2, Tim Schlaich3, Andreas Balster3, Elena Polozova4, Armin Weiser2, Annemarie Kaesboehrer2, Matthias Filter2, Hanno Friedrich3


1 Department of Population and Public Health Sciences, University of Southern California, Los Angeles, CA, USA

2 Federal Institute for Risk Assessment (BfR), 10589 Berlin, Germany

3 Kuhne Logistics University, 20457 Hamburg, Germany

4 MIT



R code

R was used to compute t-test 95% confidence intervals and create Figures 4-9 in the paper.


Germany Networks


  1. Baseline characteristic signal resonance, Germany networks: results
  • Baseline signals from simulated outbreaks: \(B^{sim}_{N_i,c,k}\)
  • Baseline signals from random sampled outbreaks: \(B^{rand}_{N_i,c,k}\)
  1. Simulated Signal Relative to Random, Germany networks, \(\overline{\psi^{sim}_{N_i}}(c)\): plot and convergence
  2. Normalized Signal Resonance measure, Germany networks, \(\overline{\Psi^{norm}_{N_i}}(c)\): results for single outbreaks
  3. Normalized Signal Resonance measure, Germany networks, \(\overline{\Psi^{norm}_{N_i}}(c)\): accuracy results on simulated outbreaks


Stylized Networks (random layered graphs, RLG)

  1. Simulated Signal Relative to Random, RLG, \(\overline{\psi^{sim}_{N_i}}(c)\): plot and convergence
  2. Normalized Signal Resonance measure, RLG, \(\overline{\Psi^{norm}_{N_i}}(c)\): accuracy results on simulated outbreaks


Instructions for using R_code

Knit the code in each markdown file index____.Rmd to reproduce analysis in each file. If the repository is built from the project file FoodItemID.Rproj, the working directory should not need to be set.


Matlab code

  • Matlab was used for all signal computations in the paper.
  • The data files (variables) for each network and outbreak are included in the file variables/workspace_variables_080521.mat
  • Before running any codes, add the functions folder and all subfolders to the working directory.
  • There are 3 categories of functions/scripts: (i) for Germany networks, (ii) for stylized networks (RLG), and (iii) global functions that are used for both.


Germany networks

There are 3 main wrapper codes that can be used to re-run the analyses in the paper. The functions used in each wrapper code can be traced back to see the full trajectory of computations from simulating an outbreak, to applying the source localization algorithm, to applying the normalized signal resonance measure, to outputting the results.

Wrapper codes

  • wrapper_BaselineSig_WHS4.m
    • Recreates baseline characteristic signal resonance plots
  • wrapper_NormSig_WHS4.m
    • Computes Normalized Signal Resonance measure \(\Psi^{norm}_{N_i,c,k}\) for a single outbreak. Results in an output matrix of \(\Psi^{norm}_{N_i,c,k}\) results of dimensions \(C\) illness intervals X \(K\) iterations
    • Specify outbreak to run in single_ob, choosing from the all_ob_data array
  • wrapper_acc_tests_WHS4.m
    • Computes, simultaneously, Simulated Signal Relative to Random] \(\overline{\psi^{sim}_{N_i}}(c)\) (see results in method0_Sig) and Normalized Signal Resonance measure \(\overline{\Psi^{Norm}_{N_i}}(c)\) (see results in method0_raw)
    • Specify number of outbreaks \(m\) to generate for each network in num_ob
    • For outbreak \(m\), specify number of iterations \(k\) to compute for each \(\Psi^{norm}_{N_i,c,k,m}\) in num_samples
    • Will compute accuracy metrics and then output files into a .csv file for processing within R
  • get_sim_ob_printPMF_WHS.m:
    • generates a single outbreak
    • applies the source localization algorithm to get pmf
    • prints the pmf


Stylized Networks (RLG)

To analyze the Normalized Signal Resonance measure on stylized networks (RLG), one must first generate a sample set of RLG, and second run the code on these networks.

  • To generate a sample set of RLG, use the script wrapper_generate_RLG.m

  • To generate a single outbreak and plot its pmf use the script get_sim_ob_RLG.m:

    • generates a single outbreak
    • applies the source localization algorithm to get pmf
    • prints the pmf

  • To recreate simulation accuracy tests on RLG, wse the wrapper script wrapper_acc_tests_RLG.m and the following instructions:

    • Specify number of outbreaks \(m\) to generate for each network in num_ob

    • For outbreak \(m\), specify number of iterations \(k\) to compute for each \(\Psi^{norm}_{N_i,c,k,m}\) in num_samples

    • If computing Simulated Signal Relative to Random (\(\overline{\psi^{sim}_{N_i}}(c)\)) for use in Normalized Signal Resonance measure, set computing_NormSig=0. If computing Normalized Signal Resonance measure (\(\overline{\Psi^{Norm}_{N_i}}(c)\)), set computing_NormSig=1.

  • Note that to compute the Normalized Signal Resonance measure, you will need to:

    1. Compute Simulated Signal Relative to Random in Matlab using the code above with computing_NormSig=0.

    2. Read computed Simulated Signal Relative to Random into R and compute mean and 95% CI statistics for each network using the code within the R markdown file, index_RLG_raw.Rmd. Save this as a csv file as e.g. RLG6_CI_4Matlab.csv

    3. Read the file RLG6_CI_4Matlab.csv into Matlab and use in the function script SimSig_L3_OB_RLG_norm.m within lines 14-17. Use the exact same num_ill_intervals as used to compute the Simulated Signal Relative to Random.



Network models in R

Please see this GitLab repository for R code for analyzing the network features of the German food supply networks, including transforming the networks into network files.