Königsweg 67
14163 Berlin
+49 30 838 56034
epi@vetmed.fu-berlin.de
Aim:
We study the geographical propagation of African Swine Fever (ASF) through various approaches. Advances in computing power, together with the amount of data obtained from disease surveillance and registries enable machine learning, standard statistical tools as well as computer simulation to be
applied to the field of Veterinary Epidemiology. Moreover, numerous mathematical models in the context of the SARS-CoV-2 pandemic show the importance of predictive models in infectious disease.
Methods:
We propose a comparison of i) statistical regression (GLM) ii) heuristic approach (quasi-gravity epidemiological SIR model solved with computer simulations) and iii) machine learning approach (XBoost). We utilised 6018 disease notifications in Poland from February 2014 to January 2020 for models training and validation. A spatial representation of notifications with a given area/region in a given time interval was built. We take into consideration 68 sequential months or 17 quartal and hexagons from 100km² to 10000km². Our ground truth is a binary state if there was registered at least one notification in a given polygon in a given time period. We have inferred our models with additional variables with precision to county/poviats (old NUTS-4) as WB (wild boars): representing natural infections between WB; pig: representing agricultural infection between farms; human: representing human mediated virus translocation due to movement of contaminated formitts or products; interactions between each layer.
Results:
Notification registry is not only showing disease propagation, but it's also sensitive to surveillance methodology, which vary in time. During the considered period national and European regulations have changed dozens of times. For example surveillance in WB was changed from passive to active (e.g. by searching for carcasses). In results much more cases can be found (with even 12 times higher test positivity rate among wild boars). Thus, five distinguished phases of epizootic were revealed: 1) Feb 2014 to Jun 2016: Sub epidemic; 2) Jul 2016 to Jun 2017: Pre epidemic; 3) Jul 2017 to Mar 2018: Epizootic - early; 4) Apr 2018 to Apr 2019: Epizootic - stable; 5) Mai 2019 to Jan 2020: Epizootic - extended.
We found that predictability of epizootic state (defined binary) one period in advance on the border between disease free and in affected regions reaches over 95% sensitivity and specificity with XBoost, which outperformed all other methods.
Conclusions:
We confirmed that ML is powerful in predicting short-term local transmission of ASF knowing simple proxies of pig farming structure, WB habitat and density of human population. Long distance jumps have been much more difficult to predict however, quasi-gravity model seems to give a good qualitative picture of the long and medium term most paths of propagation.