anty-filidor / covidout

Graph model of disease spreading

COVID OUT!

This repository contains a implementation of graph informational spreading model which simulates an illness propagation. Data is being generated in silulation of interactions between people. Internal states of people are being described by significant symptoms of CoVid19 ilness. This model has been developed after medical consultations and it is an entry for machine learning pipeline (after feeding with real data).

DOCUMENTATION IS HERE

Structure of repository

├── data_generation
│   ├── databases.example.json
│   ├── fill_database.py
│   ├── gen_database.py
│   ├── gen_graphs.py
│   ├── graphs
│   ├── names
│   ├── simulation.py
│   └── visualize_movement_simulation.py
├── docs
├── readme.md
├── requirements.txt
└── spreading_model
├── ioops.py
├── networks
├── spreading.py
├── src.ipynb
└── visualisations.py

How to run this code

Data generation

Algorithms to simulate human population inside city. There are 3 scripts to be run:

• gen_database.py: It generates database files in csv format
• fil_database.py: It fills main sql database with generated data. Please use databases.json file to enter credentials. There is databases.example.json for reference
• gen_graphs.py: It generates data for graph analisis.

Also this framework contains all the classes to parse info in database for further analysis

Spreading model

Please use src.ipynb notebook to perform an experiment. It contains

• read of network
• experiment
• saving results as json of internal states of nodes or gif:

Model is based on an observation, that its hard to distinguish between common flu and coronavirus. So instead of trying to predict coronavirus, we are tracking spread of any flu (ie. we create probabilistic chain of infections - who infected who). Then if one person from such a network is somehow diagnosed with coronaviurs, we can use this information to spread risks through is "flu symptoms" network.

To build graph structure, we are using information from the GPS / location services. Connection weights are based on time that 2 people spent in near proximity. Then to calculate risks, we use a conditional probabilities of coronavirus conditioned by symptoms (see ioops.py_comp_internal_weights) that user provided in the app and facts / probabilities of coronavirus being diagnosed in his network.

To sum up, model consists of 2 components:

• internal probability of being infected by each neighbour node. This function is a sigmoid of 13 different medical symptoms (see ioops.py_comp_internal_weights)
• external probability of illness transfer by interpersonal contact. It is stored in edges of graph model. From hat coefficients a resultant probability is being computed by using conditional probability laws.

Example

Let's assume following network, internal probability of being ill in node can be:

Weight of interactions between nodes is given as:

Thus: